Dual Adversarial Co-Learning for Multi-Domain Text Classification

Yuan Wu, Yuhong Guo

Abstract

In this paper we propose a novel dual adversarial co-learning approach for multi-domain text classification (MDTC). The approach learns shared-private networks for feature extraction and deploys dual adversarial regularizations to align features across different domains and between labeled and unlabeled data simultaneously under a discrepancy based co-learning framework, aiming to improve the classifiers' generalization capacity with the learned features. We conduct experiments on multi-domain sentiment classification datasets. The results show the proposed approach achieves the state-of-the-art MDTC performance.

Introduction

Text classification is a fundamental task in Natural Language Processing (NLP), yet it faces significant challenges due to domain shift. Language is highly context-dependent; for instance, the word "fast" may signify high quality in a sports car review but indicate poor performance in a mobile battery review. Consequently, a sentiment classifier trained on one domain (e.g., movies) often generalizes poorly to another (e.g., kitchen appliances).

Multi-Domain Text Classification (MDTC) aims to address this by leveraging knowledge from domains with abundant labeled data to improve performance in data-scarce domains. However, existing methods often suffer from "feature contamination," where domain-specific nuances leak into shared feature spaces, or they fail to effectively bridge the gap between labeled and unlabeled data within a target domain.

Methodology: The DACL Framework

The proposed Dual Adversarial Co-Learning (DACL) framework addresses these issues through a sophisticated architecture comprising three core components:

  1. Shared-Private Feature Extraction: The model employs a shared extractor ($F_s$) to identify universal patterns across all domains and a set of private extractors ($F_d$) to capture domain-specific characteristics. A separation regularizer ensures orthogonality between these feature sets, preventing the shared space from being polluted by domain-specific noise.
  2. Multinomial Adversarial Discriminator: To ensure the shared features are truly domain-invariant, a discriminator ($D$) attempts to identify the source domain of a sample. The shared extractor is trained adversarially to "fool" this discriminator.
  3. Dual Adversarial Co-Learning: This novel component utilizes two separate classifiers ($C_1$ and $C_2$). During training, these classifiers are encouraged to maximize their prediction discrepancy on unlabeled data. Simultaneously, the feature extractors are updated to minimize this discrepancy, effectively pushing unlabeled data into regions of the feature space where the decision boundary is most robust.
The DACL architecture featuring shared-private networks and dual adversarial components.

The training process follows a three-step adversarial cycle:

  • L-Step: Supervised learning on labeled data to establish baseline classification performance.
  • A-Step (Adversarial): The discriminator and the two classifiers are optimized to maximize the difficulty for the extractors.
  • R-Step (Refinement): The extractors are updated to minimize the domain discrepancy and classifier disagreement identified in the A-Step.

Experimental Setup

The DACL framework was evaluated on two benchmark MDTC datasets:

  • Amazon Review Dataset: Contains four domains (Books, DVD, Electronics, Kitchen) with 2,000 samples each, represented as 5,000-dimensional bag-of-words vectors.
  • FDU-MTL Dataset: A more diverse set comprising 16 domains (e.g., apparel, health, music). This dataset uses raw text, enabling the use of CNN-based feature extractors.

Results and Discussion

DACL achieved state-of-the-art (SOTA) results across both benchmarks. On the Amazon dataset, it reached an average accuracy of 86.59%, surpassing previous adversarial methods like MAN (85.66%). On the more complex FDU-MTL dataset, DACL achieved an average accuracy of 89.1%, outperforming the previous best model (MAN-NLL) by 0.7% and securing the top performance in 12 out of 16 domains.

Performance comparison on the FDU-MTL dataset across 16 domains, showing DACL's superior performance.

Key findings include:

  • Dual Alignment: The simultaneous alignment of different domains and labeled/unlabeled data is critical for generalization.
  • Feature Separation: Explicitly separating shared and private features prevents noise leakage, which is vital for maintaining the integrity of the shared knowledge base.
  • Robustness: DACL demonstrated strong performance even in unsupervised adaptation scenarios (zero labels for the target domain), outperforming specialized models like DANN and MDAN.

Conclusion

The DACL framework sets a new standard for Multi-Domain Text Classification by combining shared-private feature extraction with a dual adversarial co-learning strategy. While the model requires careful hyperparameter tuning for trade-off parameters, its ability to learn robust, domain-invariant features makes it a powerful tool for cross-domain NLP tasks.

Summary updated

Abstract

In this paper we propose a novel dual adversarial co-learning approach for multi-domain text classification (MDTC). The approach learns shared-private networks for feature extraction and deploys dual adversarial regularizations to align features across different domains and between labeled and unlabeled data simultaneously under a discrepancy based co-learning framework, aiming to improve the classifiers' generalization capacity with the learned features. We conduct experiments on multi-domain sentiment classification datasets. The results show the proposed approach achieves the state-of-the-art MDTC performance.
BibTeX
@inproceedings{wu2020dual,
  author = {Yuan Wu and Yuhong Guo},
  title = {{Dual Adversarial Co-Learning for Multi-Domain Text Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=SFxsCs3LEXt},
}
Figure