Tilde at WMT 2020: News Task Systems

TL;DR

This paper details Tilde's submission to the WMT2020 news translation task, focusing on English-Polish translation using Transformer models, data selection, back-translation, and ensemble techniques.

Contribution

It introduces improved translation systems with data selection and back-translation strategies, and ensemble methods for enhanced performance in news translation.

Findings

Ensembles of Transformer models improve translation quality.

Data selection and back-translation enhance system performance.

Right-to-left re-ranking contributes to better results.

Abstract

This paper describes Tilde's submission to the WMT2020 shared task on news translation for both directions of the English-Polish language pair in both the constrained and the unconstrained tracks. We follow our submissions from the previous years and build our baseline systems to be morphologically motivated sub-word unit-based Transformer base models that we train using the Marian machine translation toolkit. Additionally, we experiment with different parallel and monolingual data selection schemes, as well as sampled back-translation. Our final models are ensembles of Transformer base and Transformer big models that feature right-to-left re-ranking.