Exactly. And all the inefficiencies that come from multiple "target" languages. What happens when the Indinesian scientist decides to learn German and Chinese, while the Latvian scientist learns Spanish and English. They've both taken the effort to learn 2 languages and yet still won't be able read each other's papers. It's just simpler to have one target where you know everyone else is aiming too.

Obviously we need an intermediate representation language, which we could hypothetically call "Language and Lexicon for Vernacular Metatranslation". From this language, we could generate automatic translators to and from various target languages, including English, French, and so on. What would be useful is that we could then write reusable prose tightening tools in the Language and Lexicon for Vernacular Metatranslation that could be shared across all the target languages.

Scientists could write their science in any of the higher level representation languages, or even Esperanto or Lojban, or whatever they culturally prefer for their particular brand of science. As a bonus, you could even use the Vernacular Metatranslator optimizers to tighten up prose even if it was originally written in your preferred output language!

Except, unfortunately, natural languages don't really work this way. There's so much additional information spliced in phrases, constructs, choice of words, tone, etc. that building this Vernacular Metatranslator is an GAI-complete task. At this point we're back in the Babelfish territory. Not saying it's impossible, but we seem to be far from required technology level to do that.

But you know, this is how the recent work on deep networks for translating work and why they deliver such nice results: you have two recurrent neural networks, joined together in the middle.

You mean human brains, of course?

No? I mean deep neural networks.

Sorry. I wasn't sure if "two recurrent neural networks, joined together in the middle" was something in DNN research (I'm not very up to date with the current state of it) or an euphemism for human brain (sarcasm sometimes comes in very subtle forms here on HN).

Something like that exists: http://en.wikipedia.org/wiki/Esperanto_grammar#Grammatical_s...

In most languages, word for word translation produces rubbish. In Esperanto it doesn't.

Esperanto has no particular word order, instead, suffixes denotes parts of speech. This alone makes it much more simple to translate literally between languages. It also has no gender, and is perfectly regular.

The language was designed so that one could send a letter to a friend who with did not speak it, and they could translate the entire document quickly using only a dictionary and a basic understanding of the grammar.

> But what's the alternative?

Training the whole world in an artificial language like Esperanto. On one hand, this has a cost now even English speaking people have to learn a second language, on the other hand an artificial language is MUCH easier to learn than our 'natural' languages with all their inconsistencies..