Video Recordings

Machine translation (MT) is still evolving; we can see many new species coming from related areas (machine learning) and a revival of endangered species (linguistically adequate MT). The talk will start with a brief taxonomy of MT systems with a special focus on dependency-tree-based MT. I will describe some machine learning techniques from which we can benefit in machine translation, such as large-scale discriminative training and online algorithms for structured prediction. The second part of the talk will report on our ongoing effort to improve TectoMT - a deep-syntactic tree-to-tree MT system. I will present a novel decoding algorithm and a related training approach based on guided learning which should substitute the current transfer phase of TectoMT. I will also mention two hybrid approaches that combine TectoMT with standard phrase-based MT (Moses).