Just guessing, but I imagine that Google Translate probably will, assuming, say, we represent Zapotec to English as f and English to Kannada as g use f ∘ g to i

Just guessing, but I imagine that Google Translate probably will, assuming, say, we represent Zapotec to English as f and English to Kannada as g use f ∘ g to infer the translation of Zapotec to Kannada. And saying that, I’m guessing that if the corpus of f⁻¹ is too small, it can probably attempt to infer the gaps from f. But these inferred mappings are probably of lower quality than explicit ones returning us to the original issue. And man, I really wish someone who actually knew would weigh in on this.