Theiling Online    Sitemap    Conlang Mailing List HQ   

Re: Automated translation

From:Raymond A. Brown <raybrown@...>
Date:Saturday, March 13, 1999, 8:15
At 11:22 pm -0500 12/3/99, Carlos Thompson wrote:
>Herman Miller wrote: > >> Does anyone here have any familiarity with writing software to do automated >> translation of languages? > >[...] > >> . Has anyone done any >> experimentation in this area, either with their conlangs or natlangs?
The dissertation for my master's degree in computing was concerned with just this. As part of it I did produce an automatic translator, written in Prolog, which translated a subset of German into English. The English was always readable in the way that some automatic translators are not - but would need 'touching up' to produce 100% natural English; e.g. the preterite was simply translated 'did' + infin. It was then up to the English editor to get the best English past tense for the context.
>I'm suposed to be working in a project on automatic translation... well, the >project includes speech recognition, automated translation and sign synthesis. >(I'm focused on the translation part.)
Wow!! Now speech recognition does have problems. It'll be interesting to know how the project gets on.
> >An small summary of what we have found: there is not yet a perfect translating >machine.
Too right - and some are pretty hopeless. Not so long ago I found an interesting web-site in French and clicked the button for 'translation' to English when I printed it. Oh dear - the result was almost gibberish. I reprinted the French - it was far more intelligible.
>There are some heuristics models and one interesting I heared about >is a big AI machine (probably a neural machine), feed it with the phrases in >the source language, feed it with translations in the destination >language, and >let the machine find patterns. Then it could translate any phrase... the >problem is the amout of phrases you have to feed the machine with.
Sounds interesting - do you know how successful the system is?
>There is the modular alternative we are following: 1. try to understand >what is >said in the source language; 2. translate the structures and the >vocabulary; 3. >generate the correct forms in the destination language. > >For understanding what is said, there is a morphological analysis and a >syntactic analysis. Then you have a tree with all the analisys of each >sentence. Finnaly you have to take ambiguity away, and there is no real >method... if "I saw the wood yesterday", you have the clues: yesterday -> past >-> saw = see_past... but this is the most difficoult part and you say is not >your intention to come that close.
Yep - that's precisely the way I went about things on my small model.
>Those trees can be converted, some branches moved (the past decletion in a >verb >would become a particle and the definite article would become a shift in the >order of the sentense, and the noun phrase before the verb would be given a >nominative ending... et cetera). > >Once you have the tree you would generate, first syntacticaly, then >morphologically, and you have translated.
Yep - that's about it. Moving those branches around is quite challenging and, at least for me, interesting. ........
> >The kind of output makes me thing, that the best way of manipulating the trees >are languages like LISP (but I had no echo at work... probably because nobody >in the group, including me, knew anything on Lisp).
Yes, for those who know & have access to LISP. Personally, it's a language I never took to. Prolog is also very good at manipulating trees. ........
> >Well, it seems not to be imposible.
Certainly, more pedestrian language is probably susceptible to some sort of satisfactory automated translation, but poetry?? Who knows :) It's a fascinating area of research and if it weren't for that awkwark business of having to earn a living, I'd like to get back in there again. My advice to Herman is to go for it and try it. Ray.