12

More than fifty years ago, philosopher Yehoshua Bar-Hillel wrote wrote an influential paper about computerized translation entitled: A Demonstration of the Nonfeasibility of Fully Automatic High Quality Translation (see also this TIME magazine article from 1954). Bar Hillel wrote several further papers on the matter.

Here are some related links: A 2000 paper entitled Yehoshua Bar-Hillel: A philosopher's contribution to machine translation by John Hutchins; the paper Bar Hillel and Machine Translation: Then and Now by Sergei Nierenberg, and the Wikipedia articles on Bar Hillel and machine translation.

My questions are: to what extent are Bar-Hillel's comments still relevant? How do they fit with the experience we have had over the last half century?

Improve this question
2
  • I removed the following text from the end of the question: "Are there further foundational philosophical works on machine translation in general and new arguments/insights regarding the difficulty of machine translation?" We encourage questions to be self-contained and focused on a single point. That text would make a great follow up question(s), perhaps after this one has accumulated some answers so that people can reference and expand on them.
    – Aaron
    Sep 28, 2011 at 20:57
  • Thanks for the edit, Aaron. In case somebody will want to follow up on that, the additional questions were: Are there more recent writings by philosophers on machine translation? and are there new arguments or insights regarding the difficulty (or even infeasibility) of high-quality machine translation?
    – Gil Kalai
    Sep 30, 2011 at 19:00

2 Answers 2

Reset to default
5

The Bar-Hillel article deals mostly with the ability to disambiguate words in isolation and with the aid of context. While an ambiguous word in isolation is difficult or impossible even for humans (e.g. which word am I referring to when I say "pool"?), a machine could at least make an attempt to choose the word-sense with the highest probability of occurring. Even this though needs to be trained on a wide range of domains in order to get the most accurate statistical range for the word in question.

As far as determining a word-sense from context, I would say that the article is severely outdated, as there are many algorithms that can achieve very accurate results. One of the methods involves looking at sources like WordNet to determine the likelihood of a word sense based on the maximum likelihood estimate of the other non-function words in the context range.

There are many articles describing these methods and more available online. Perhaps some of these sources will be useful: Google Scholar - "Machine Disambiguation"

Improve this answer
5

I think the claim of Bar-Hillel is two-fold that, at that time the two approaches, which I will call roughly n-gram (or fixed length context) and thesaurus methods, were doomed to failure – as he projected at that time.

I think the n-gram approach has been vindicated since then. I think the reasons are not because of better use or manipulation of contexts but because of simply:

  • better corpora (masses of working data)
  • better processing capability (simply speed)
  • restriction to local translation contexts (technical documents as opposed to literature).

The thesaurus method (which is really nowadays an 'ontology' method) has not been incorporated well into MT (as far as I know so far), but I expect that the same 3 situations (better corpora, processing capability, restricted contexts) will make MT even more successful.

That is, I think the ideas mentioned by Bar-Hillel as failures of the 1960's MT technologies are just not failures any more, but strengths. I expect soon enough that those methods will not be enough, but over the past decades, MT really has gotten much much better. No, it's not totally hands-off automatic, but then neither is driving.

(disclaimer: I had never read the Bar-Hillel paper before now. Also, I am not an MT researcher or implementer)

Improve this answer

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.