2020/04/27 by M Saiful Bari, Bari, M Saiful, Tasnim Mohiuddin +3
Computer Science · #Computation and Language (cs.CL) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Multimodal Machine Learning Applications #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2004.13240
openalex publication_date 2020/04/27 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Transfer learning has yielded state-of-the-art (SoTA) results in many\nsupervised NLP tasks. However, annotated data for every target task in every\ntarget language is rare, especially for low-resource languages. We propose\nUXLA, a novel unsupervised data augmentation framework for zero-resource\ntransfer learning scenarios. In particular, UXLA aims to solve cross-lingual\nadaptation problems from a source language task distribution to an unknown\ntarget language task distribution, assuming no training label in the target\nlanguage. At its core, UXLA performs simultaneous self-training with data\naugmentation and unsupervised sample selection. To show its effectiveness, we\nconduct extensive experiments on three diverse zero-resource cross-lingual\ntransfer tasks. UXLA achieves SoTA results in all the tasks, outperforming the\nbaselines by a good margin. With an in-depth framework dissection, we\ndemonstrate the cumulative contributions of different components to its\nsuccess.\n