Andre,

I’d be interested to know if transformer based architectures can solve the problem of implicit language translation without explicitly learning n^2 different pairs of languages. LSTM can encode language into an abstract representation (what professor Hinton at some point called a “thought vector”) and then if you already have English-German and English-Chinese, getting German-Chinese should be an easier task.

I think the transformer might be modified to solve this problem, but I’m not sure if (which one of) the existing architectures do that?