4 ms·

Obviously we need an intermediate representation language, which we could hypothetically call "Language and Lexicon for Vernacular Metatranslation". From this l

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!

TeMPOraL 12y ago
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.
- gwern 12y ago
  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.
  - TeMPOraL 12y ago
    You mean human brains, of course?
    - gwern 12y ago
      No? I mean deep neural networks.
      - TeMPOraL 12y ago
        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).
mrrrgn 12y ago
Something like that exists: http://en.wikipedia.org/wiki/Esperanto_grammar#Grammatical_summary 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.

TeMPOraL 12y ago

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.

gwern 12y ago
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.
- TeMPOraL 12y ago
  You mean human brains, of course?
  - gwern 12y ago
    No? I mean deep neural networks.
    - TeMPOraL 12y ago
      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).

mrrrgn 12y ago

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