Google says its Neural Machine Translation System translates between language pairs it hasn't seen before, and sees signs of “interlingua” common representation
Posted by Mike Schuster (Google Brain Team), Melvin Johnson (Google Translate) and Nikhil Thorat (Google Brain Team)
Context & Ripple Effects
This closes out a rapid 2016 arc for Google Translate: neural machine translation first shipped for Chinese-to-English conversion, then landed in the Translate app across eight languages with full-sentence translation just ten days before this post. What the Google Brain and Translate teams add here is the structural claim — the system translates between language pairs it was never trained on, which they read as evidence of an "interlingua" common internal representation.
First-order effects
- Google no longer needs a separately trained model per language pair: pairs absent from training data become translatable the moment both languages exist anywhere in the network, collapsing the marginal cost of adding coverage.
- The finding validates the eight-language app rollout's architecture, giving the Google Brain Team a research result that doubles as justification for rebuilding Translate on one shared system.
Second-order effects
- Coverage expansion becomes an engineering decision rather than a data-collection project — consistent with the follow-on extension of the AI-powered system to Russian, Hindi, and Vietnamese, where shared representations let Google scale to more languages on the same infrastructure.
- Rival translation providers are pushed toward multilingual architectures too, since a per-pair competitor now faces a cost curve that improves with every language Google adds instead of degrading.
Third-order effects
- If intermediate text is only an artifact of current interfaces, the endpoint is removing it entirely — a path Google itself later traced with Translatotron's direct speech-to-speech translation, which skips text as the pivot between languages.
- Shared-representation translation points toward translation dissolving into general-purpose multilingual models, making language support a property of the model rather than a product feature shipped pair by pair.
The trend: Machine translation is shifting from independently trained per-pair models to shared multilingual representations that decouple language coverage from training data — with direct speech-to-speech as the logical terminus.