Is Google translate's current NMT system able to take into account higher-than-sentence-level context when translating?

[Russell Willoughby]

I am currently a master's student at the University of Geneva, completing a thesis on machine translation. My research experiment includes the use of google translate. I  have a question regarding google translate's NMT architecture that I was hoping someone would be able to help me with. Is Google translate's current NMT system able to take into account higher-than-sentence-level context when translating?

 I have noticed that I sometimes get different results when I feed one sentence into google translate, for example, versus when I feed that same sentence + the surrounding paragraph into google translate...However, it was my understanding the NMT systems were only able to "understand" text on a sentence-to-sentence basis. Is this the case, or is google translate able to take into account some context when translating?

Thank you so much!

[Michel Farhi]

Interesting question, would like to hear what Google says.   I can tell you that DeepL has an awareness of context beyond a single sentence.  For example, test this text "My car is new.  It has four wheels. "   into French.  You will see that DeepL trannslates "It" into "Elle" in French, as it recognizes that "it" refers to car, which is feminine in French.  Until recently, my test of the same paragraph in Google translate was getting it wrong.

This said, when you translate with a TMS or a CAT tool, they typically segment sentence by sentence, and thus will not even attempt to send mutiple sentences to the MT engine for a grouped translation.  That's a limitation of CAT tools, not the MT engine.

Good luck with the thesis.   What's your topic?