Multilingual Translator

A training method for multilingual neural machine translation systems that is efficiently extendable to new languages and data modalities. Training neural machine translation systems is challenging both in terms of data and computational resources. These factors become more critical in the multilingual setting, where several languages may be input or generated by the system. The most widespread strategy consists of training a single sequence- to-sequence system shared between all languages. This architecture allows knowledge transfer but it forces a dependency between languages. This dependency limits the ability of the system to efficiently extend to new languages or modalities. The whole system has to be retrained using data for all languages which implies variations in the overall performance. This technology based on the sequence-to-sequence architecture for neural machine translation defines how to efficiently train a multilingual system extendable to new languages and modalities while allowing knowledge transfer. The process consists of two main steps. Firstly, joint training of languages specific encoder and decoders to a common language representation without parameter sharing. Secondly, incremental training of new languages and modalities to the system, including an additional module to mitigate the differences between speech and text representations.