Stochastic Adversarial Networks for Multi-Domain Text Classification

Xu Wang, Yuan Wu

Keywords: Multi-domain text classification, Adversarial training, Multivariate gaussian distribution

TL;DR: We propose a stochastic adversarial network to reduce the required model parameters for multi-domain text classification.

Abstract: Adversarial training has played a pivotal role in the significant advancements of multi-domain text classification (MDTC). Recent MDTC methods often adopt the shared-private paradigm, wherein a shared feature extractor captures domain-invariant knowledge, while private feature extractors per domain extract domain-dependent knowledge. These approaches have demonstrated state-of-the-art performance. However, a major challenge remains: the exponential increase in model parameters as new domains emerge. To address this challenge, we propose the Stochastic Adversarial Network (SAN), which models multiple domain-specific feature extractors as a multivariate Gaussian distribution rather than weight vectors. With SAN, we can sample as many domain-specific feature extractors as necessary without drastically increasing the number of model parameters. Consequently, the model size of SAN remains comparable to having a single domain-specific feature extractor when data from multiple domains. Additionally, we incorporate domain label smoothing and robust pseudo-label regularization techniques to enhance the stability of the adversarial training and improve feature discriminability, respectively. The evaluations conducted on two prominent MDTC benchmarks validate the competitiveness of our proposed SAN method against state-of-the-art approaches.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7001