My recent research interests include discourse-oriented machine translation. One of the main shortcomings in current systems of automatic translation is the focus on local phenomena and the translations of sentences in isolation.

It seems obvious from a linguistic point of view that discourse-level relations should be considered to produce cohesive and coherent translations. My former PhD student, Christian Hardmeier, initiated a new line of research in which document-wide dependencies can be integrated in statistical machine translation. One of the frustrating results of his research was that success in this field is hard to achieve and it seemed necessary to have a larger research community to make any progress.

Fortunately, several other groups became interested in this topic and together we initiated the workshops on discourse in machine translation with the catchy name DiscoMT. The latest edition of the workshop also featured a shared task on machine translation with a focus on referential pronouns. The workshop was certainly a success, however the shared task once again revealed the difficulties of improving current systems with discourse-related features. This intriguing challenge requires further investigation that keeps us busy. Research is fun even with negative results.