Quick and painless domain adaptation for SMT

Date: 
14/03/2012
Heure: 
10h30
Orateur: 
Rico Sennrich
Laboratoire: 
Institute of Computational Linguistics, University of Zurich
Résumé du séminaire: 
As more and more parallel data for SMT training becomes available, domain adaptation techniques are required to prevent relevant in-domain data from being diluted. I present results and ongoing work on domain adaptation for SMT, focusing on translation models and demonstrating techniques that require little supervision and are suitable for large-scale systems. The overview touches on how one can perform domain adaptation with 100 corpora, and how to treat system combination as a domain adaptation problem.