[Nettime-bold] Re: <nettime> The cultural bias of translating programs

[Keith Sanborn <mrzero@panix.com>]

I think there's a bit more involved than dictionary tweaking in
translation programs, although it does explain the French/English
vs. Portuguese/English example fairly well. The problem, of course, is
what one understands by a dictionary, which as Julian Dibbel is highly
labor intensive.

What there is more to it than dictionary tweaking is the level of
complexity of natural languages, limitations in computing power, and the
fact that languages mutate in ways that may not be able to be accounted
for for some time if ever. In a way, it's still like a bunch of very
gifted monkeys (the tweakers) with typewriters. How many would it take how
long to produce a decent "dictionary" i.e. a set of fairly reliable
correlations between two natural languages and a set of algorithms for
parsing texts in those languages which would allow the dictionaries to be
applied. I think the rules for parsing natural languages are
extraordinarily complex.

One might argue that the level of artificial stupidity produced by
translation programs is a good measure of just how distant we are from
making an adequate machine based account of human intelligence.

In addition, it is very difficult to account for the cultural history
built into a natural language which is what makes dictionary reading so
wonderful in the first place.

Finally, our dictionary tweakers are very much in the position currently
of Flaubert's Bouvard and Pecuchet. No matter how much they copy and
transcribe they will never render anything but an aburdly literal account
of things. And evern if they were successful temporarily, unless their
parsing engines were extremely intelligent--in fact self-modifying--they
would be, as Saussure described Esperanto--like a duck trying to hatch a
hen's egg. You can never step into the same river of language twice.

Keith Sanborn



On Tue, 14 Nov 2000, Julian Dibbell wrote:

> Felix Stalder, commenting on Wade Tillett, wrote:
> 
> > >I wondered how much text degrades while it is put in a translator. Like
> > >the experiment, I wrote this text in the babelfish and translated it the
> > >French-English one and the English-French one until the text becomes '
> > >produces ' the left side with us see...
> >
> <...>
> > What was really interesting in Wade's experiment is to see that a text
> > indeed does stabilize. Stabilization indicates that this version of the
> > text contains only words that are, from the point of view of the
> > translating program, unambiguous in both languages. And the way it
> > stabilizes reveals the bias of the translating algorithm. In Wade's case,
> > the translator seems reveals a bias towards business prose (probably
> > that's where the market is for the high-end version).  It would be
> > interesting to see if stabilization occurs in all translating programs at
> > the same point, or if different translating programs have different types
> > of biases.
> 
> 
> See also my "After Babelfish" (at
> http://www.feedmag.com/book2/essay_dibbell.html and also posted to nettime a
> few weeks back). In that essay I did an experiment similar to Wade's, using
> a Yeats poem and the Portuguese-English Babelfish. That program tends to
> produce a slightly wackier text than those that translate French or German,
> I find. Why? Because Portuguese is a "minor" language, with a smaller
> market, and therefore has gotten less attention from the Babelfish
> dictionary tweakers. And there's no getting around it: only many, many
> person-hours of hand-tweaking a given language-pair can produce a halfway
> decent machine translator. There is no magic algorithm for it. Consequently
> these are very capital-intensive -- and therefore market-sensitive --
> products. Which means that the English-French translator does a better job
> than the English-Portuguese, and the Icelandic-Hindi will never see the
> light of day.
> 
> So Felix is right in more ways than one. Markets do help build cultural bias
> into machine translators -- not only by warping the vocabulary of a given
> language pair toward a particular context (e.g., the business world), but by
> warping the overall effectiveness of the pair as well.
> 
> #  distributed via <nettime>: no commercial use without permission
> #  <nettime> is a moderated mailing list for net criticism,
> #  collaborative text filtering and cultural politics of the nets
> #  more info: majordomo@bbs.thing.net and "info nettime-l" in the msg body
> #  archive: http://www.nettime.org contact: nettime@bbs.thing.net
> 


_______________________________________________
Nettime-bold mailing list
Nettime-bold@nettime.org
http://www.nettime.org/cgi-bin/mailman/listinfo/nettime-bold