The LMU Munich System for the WMT 2020 Unsupervised Machine Translation Shared Task

This paper describes the submission of LMU Munich to the WMT 2020\nunsupervised shared task, in two language directions, German<->Upper Sorbian.\nOur core unsupervised neural machine translation (UNMT) system follows the\nstrategy of Chronopoulou et al. (2020), using a monolingual pretrained language\ngeneration model (on German) and fine-tuning it on both German and Upper\nSorbian, before initializing a UNMT model, which is trained with online\nbacktranslation. Pseudo-parallel data obtained from an unsupervised statistical\nmachine translation (USMT) system is used to fine-tune the UNMT model. We also\napply BPE-Dropout to the low resource (Upper Sorbian) data to obtain a more\nrobust system. We additionally experiment with residual adapters and find them\nuseful in the Upper Sorbian->German direction. We explore sampling during\nbacktranslation and curriculum learning to use SMT translations in a more\nprincipled way. Finally, we ensemble our best-performing systems and reach a\nBLEU score of 32.4 on German->Upper Sorbian and 35.2 on Upper Sorbian->German.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC