Boosting Unsupervised Machine Translation with Pseudo-Parallel Data

• DOI: 10.48550/arxiv.2310.14262
• Publication Date: 2023-01-01
• Authors: Ivana Kvapilíková; Ondřej Bojar
• Keywords: Boosting; Parallel corpora; BLEU; Baseline (sea)

Even with the latest developments in deep learning and large-scale language modeling, task of machine translation (MT) low-resource languages remains a challenge. Neural MT systems can be trained an unsupervised way without any resources but quality lags behind, especially truly conditions. We propose training strategy that relies on pseudo-parallel sentence pairs mined from monolingual corpora addition to synthetic back-translated corpora. experiment different schedules reach improvement up 14.5 BLEU points (English Ukrainian) over baseline data only.