Automatic translation from one language to another using machines, aka machine translation (MT), has been one of the main goals of AI. The majority of the MT literature works at the sentencelevel, by treating the document to be translated as a bag of sentences and translating sentences independently. Coherent translation of documents is out of reach for current state-of-the-art MT systems, since several discourse phenomena cannot be translated correctly without referring to extra-sentential context. My work involves capturing the wider extra-sentential context in MT, particularly when translating documents. Incorporating sentence semantics into MT is largely overlooked by the research community, due to the modeling and algorithmic challenges. My ongoing work integrates the semantic structure of a sentence into the translation process, replacing memorisation with generalisation. Relevant topics are memory networks and neural Turing machines, in order to learn powerful text-to-meaning and meaning-to-text transducers from rich spaces of functions.