It seems that most projects attempt to teach the AI to learn individual, specific languages.

It occurs to me that there are relations in written and spoken words and phrases across languages - most of use have a much easier time learning more languages after we learn a second language, and we start to understand the relations between words and phrases in different languages.

Has anyone attempt to train an AI to learn all languages?

Wouldn't this potentially be a much simpler problem than trying to teach an AI a single, specific language with all of the specifics and details of that single language? Since you're actually omitting a lot of related data in other languages from the training set?

  • $\begingroup$ One basic problem is the question "what does it mean to teach a computer system a language?" Language is an instrument for communication, so there would have to be something to communicate, which is not necessarily a trivial issue. Suppose you have a system that "knows" a language, what would it do with it? $\endgroup$ Sep 4, 2017 at 15:34

1 Answer 1


There are approaches in machine translation that try to capture this kind of synergy between languages. The idea is that if you train your architecture to be able to translate English-Japanese, Japanese-English, Korean-English, English-Korean it will also be able to translate from Japanese to Korean without ever having seen a single such training example. Here, you can read about this so-called zero-shot translation.

It is also possible to train wordvectors on several languages at once, which might give you better wordvectors for a language with few training examples.

Of course for true language understanding you have to solve the grounding problem and I'm not sure using several language will help with that.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .