Wasserstein GAN based on Autoencoder with back-translation for cross-lingual embedding mappings

摘要

Recent works about learning cross-lingual word mappings (CWMs) focus on relaxing the requirement of bilingual signals through generative adversarial networks (GANs). GANs based models intend to enforce source embedding space to align target embedding space. However, existing GANs based models cannot exploit the underlying information of target-side for an alignment standard in the training, which may lead to some suboptimal results of CWMs. To address this problem, we propose a novel method, named Wasserstein GAN based on autoencoder with back-translation (ABWGAN) that can effectively exploit the target-side information and improve the performance of GANs based models. ABWGAN is an innovative combination of preliminary mappings learning and back-translation with target-side (BT-TS). In the proposed BT-TS, we back-translate target-side embeddings with preliminary CWMs to learn the final cross-lingual mappings, which enables to improve the quality of the preliminary mappings by reusing the target-side samples. Experimental results on three language pairs demonstrate the effectiveness of the proposed ABWGAN. (C) 2019 Elsevier B.V. All rights reserved.