Toward Domain Translation with Monolingual Domain Data Only
Keywords: Neural Machine Translation, Unsupervised Domain Adaptation, Energy-Based Models, Conditional Distributional Policy Gradients
TL;DR: We address unsupervised domain adaptation in NMT without using parallel data. This is achieved by utilizing the distribution of target-side monolingual data through energy-based models (EBMs) and Conditional Distributional Policy Gradients (CDPG).
Abstract: Neural machine translation (NMT) is very sensitive to domain shifts requiring a carefully designed fine-tuning strategy to avoid catastrophic forgetting problems when adapting to a new domain. Fine-tuning usually relies on high quality in-domain data, but constructing a sufficient amount of parallel data for training poses challenges even for fine-tuning. In contrast, domain-specific monolingual resources are more accessible when compared with bilingual data. Therefore, we challenge the domain adaptation of a general NMT model using only features obtained from a small amount of monolingual data. We regard the task as an instance of domain shifts, and adopt energy-based models (EBMs) and approximate these EBMs using Conditional Distributional Policy Gradients (CDPG). Recent work has applied CDPG with a small number of EBMs for NMT models limiting the capacity for domain shifts, but we construct a large number of EBMs considering the entire domain-specific data, i.e., unigram distribution, and perform fine-tuning according to their constraints. Our results show that fine-tuning using a large number of EBMs can achieve a robust domain shift without causing catastrophic forgetting, demonstrating a robust domain shift using only a small amount of monolingual resources.
Primary Area: applications to computer vision, audio, language, and other modalities
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 13869
Loading