Dr. Mirela-Stefania Duma

I can be reached the following way:
* Email: mduma at informatik.uni-hamburg.de
* Room: F-405
* Telephone: 42883-2322
* Group: NatsGroup

I am a PhD student at the University of Hamburg since October 2013 and I joined the Natural Language System Division (NATS) in January 2013. I am interested in Machine translation, Domain adaptation, Semantic Web, Information retrieval.

My thesis focused on performing domain adaptation for SMT. The common assumption made when evaluating a system is that the test distribution and the training distribution are drawn from the same data. However, in real-life applications a system is trained on a general domain and used on a specific domain. This leads to poor performance, thus the challenge is to create applications that can be robust to changes. With the increasing amount of data being available, there is a need to choose what information to use in training a system. In the case of a statistical machine translation system it has been proven that training on more data gives better results. However, this implies training on a general domain and in the case of machine translation (as in the case of other tasks like speech recognition or opinion mining) the evaluation is done on specific domains that sometimes can be highly divergent from the general domain. The divergence comes from syntactic, semantic and pragmatic differences between domains. This leads to the need to perform domain adaptation in machine translation.