Domain adaptation in statistical machine translation systems via user feedback

Machine translation research has progressed in recent years thanks to statistical machine learning methods, sufficient computational power, open source tools and increasing availability of bilingual parallel text resources. However, most of these systems stay in the hands of researchers and are not improved with public users in mind The motivation behind this thesis is the vision of freely available machine translation systems. They may be particularly important for languages and domains where there is not enough commercial interest for providing such services otherwise. The main focus of this work was to collect reference translations for Finnish news sentences, and to use this data to improve a baseline translation system on this news domain. A web application was created for rating and correcting translations and volunteers were invited to participate the effort. Then, three different approaches to domain adaptation were realized and evaluated using the news domain data. In particular, language and translation model interpolation and post-editing have been studied. Thanks to volunteers, a 1000 sentence bilingual Finnish-English news corpus was assembled. The corpus is a good asset for further research in domain adaptation. The adaptation results show that a combination of language model and translation model interpolation effectively adapts the baseline system to the news domain. Using available domain adaptation methods, translation systems can be built with simple means and adjusted to the users' needs by community feedback