UvA

The central topic of this thesis is the exploration of properties that are common across languages and properties that differentiate them, in the context of machine translation. The field of machine translation aims to render the content expressed in one language into another language. Therefore, the difficulty of this task is proportional to how different the given two languages are in expressing the same information. In the first part of the thesis, we focus on properties common across languages. We aim to validate the hypothesis that every language has a level of representation which is shared, to some extent, by all languages. This level of representation is the syntactic structure of a sentence. Consequently, one can narrow down the search for translation correspondences between units of any two languages. We realize these ideas in the form of bilingual syntactic language models which are used as soft constraints during the translation process. The models provide improvements in translation quality in a series of rescoring and decoding experiments. In the second part of the thesis, we use the idea that despite the observation that all natural languages share certain ways of expressing content, there are also many features that differentiate them. We hypothesize that the differences between languages are systematic. We design ensembles of neural machine translation systems sharing the target language but differing in their source languages. We perform a comparison to monolingual ensembles obtained by initializing systems with different random seeds and observe systematically better performance for the multi-source ensembles.