SYSTRAN Purely Neural MT Engines for WMT2017

TL;DR

This paper details SYSTRAN's neural machine translation systems for WMT2017, utilizing OpenNMT with LSTM and attention, enhanced by back-translation and test sentence adaptation for improved English-German translation.

Contribution

Introduces neural MT systems with monolingual data augmentation and test-specific fine-tuning for WMT2017 translation tasks.

Findings

Effective use of back-translation improved translation quality.

Test sentence adaptation enhanced model performance.

Systems achieved competitive results in WMT2017.

Abstract

This paper describes SYSTRAN's systems submitted to the WMT 2017 shared news translation task for English-German, in both translation directions. Our systems are built using OpenNMT, an open-source neural machine translation system, implementing sequence-to-sequence models with LSTM encoder/decoders and attention. We experimented using monolingual data automatically back-translated. Our resulting models are further hyper-specialised with an adaptation technique that finely tunes models according to the evaluation test sentences.