LIG statistical machine translation systems for IWSLT 2010

This paper describes the systems developed by the LIG laboratory for the 2010 IWSLT evaluation. We participated to the AE BTEC task and to the new TALK task. For AE BTEC task we developed two different systems: a statistical phrase-based system and a hierarchical phrasebased system using the Moses toolkit. The combination of these systems, which improves the results on different development sets, makes our final submission. This year, we concentrated on the new TALK task. The development of a reference translation system, as well as an ASR output translation system, is presented. For this latter task, re-punctuating the ASR output, before translation, seems to be very useful, while segmenting the ASR flow, which is also discussed in this paper, has shown to be less useful. Unsuccessful attempts to exploit ASR lattices instead of ASR 1best are also presented at the end of this article.