Rich Contexts in Neural Machine Translation

Machine translation has been dominated by phrase-based statistical machine translation (SMT) for the last decade, even though SMT comes with some well known limitations regarding the performance of these systems. For instance, phrase-based models need to make strong independence assumptions and only consider local context during translation, which makes it hard to model phenomena such as long-distance dependencies within sentences.
Recently, a new approach to machine translation systems has received considerable attention: machine translation with neural networks. Even though artificial neural networks have been studied for many years in other disciplines, they are a new and promising area of research within machine translation.

could not display image

In this project, we work on end-to-end neural machine translation (NMT). We explore the opportunities of neural MT with respect to the inclusion of additional information into the translation model, such as document-level information, linguistic annotation, multiple source languages, or side constraints.

With this approach, we explore the new possibilities of neural machine translation to advance state-of-the-art MT systems for the three major languages of Switzerland: French, German and Italian.

Project Head

Martin Volk


Rico Sennrich

Annette Rios

Mathias Müller, M.A.

This project is funded by the Swiss National Science Foundation under grant 105212_169888 and starts in January 2017.