Abstract
The development of architectures exploiting " deep " neural learning methods in machine translation has led to a considerable increase in the acceptability and usability of machine-computed translations. These new architectures have also made it possible to implement machine translation devices that go beyond the usual framework of translating a source-language text into a target-language text : direct translation of speech, joint translation of text and image, and so on. In this talk, I will present one such device, designed to translate from multiple source languages into multiple source languages, highlighting the computational and linguistic benefits that such multilingual translation systems bring, particularly for translating from and into minority languages.