The remarkable success of end-to-end machine translation (Sutskever et al 2014) calls into question every stage of the traditional analysis -- transfer -- synthesis MT pipeline, including the very idea of a dictionary that has before been considered pivotal for the transfer stage. Yet dictionaries remain relevant both for human language learning and proficiency and, as we shall argue here, both as an abstract conceptual representation, exactly the kind that many NLP researchers interested in representation learning urge (Bengio 2013, Scholkopf et al, in press) and as a downstream task assessing the quality of the e2e system.

Slides Slides are in English, the lecture was in Hungarian, see Video