Beyond multi-class – structured learning for machine translation

In this thesis, we explore and present machine learning (ML) approaches to a particularly challenging research area – machine translation (MT). The study aims at replacing or developing each component in the MT system with an appropriate discriminative model, where the ultimate goal is to create a powerful MT system with cutting-edge ML techniques.The study regards each sub-problem encountered in the MT field as a classification or regression problem. To model specific mappings in MT tasks, the modern machine learning paradigm known as “structured learning” is pursued. This approach goes beyond classic multiclass pattern classification and explicitly models certain dependencies in the target domain.Different algorithmic variants are then proposed for constructing the ML-based MT systems: the first application is a kernel-based MT system, that projects both input and output into a very high-dimensional linguistic feature space and makes use of the maximum margin regression (MMR) technique to learn the relations between input and output. It is amongst the first MT systems that work with pure ML techniques. The second application is the proposal of a max-margin structure (MMS) approach to phrase translation probability modelling in an MT system. The architecture of this approach is shown to capture structural aspects of the problem domains, leading to demonstrable performance improvements on machine translation. Finally the thesis describes the development of a phrase reordering model for machine translation, where we have compared different ML methods and discovered a particularly efficient paradigm to solve this problem