[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [aymara] Aymara and Western Culture

> From: "Laura Jones" <lmjones@ptd.net>

> Hi Jorge!  This is tremendously interesting - I bet that quote led to the
> first mention I saw of the perfection of Aymara.  

Dear Ms. Jones,

With respect, I must urge caution about talk of "perfection" in
natural languages, whether for Greek, Latin, Aymara or anything else.
To a professional linguist, all languages are potentially
interesting, and none is more "perfect" than another.

One used to hear a lot about the supposed perfection of ancient
Greek and Latin, and about the "degeneration" of their descendents,
from old classics masters.  Such claims are rejected by all trained 
linguists today, not from political correctness but because they 
simply do not hold up under any objective criteria.

As for claims for the perfection of Aymara, coming notably from
Iván Guzmán de Rojas and picked up by Umberto Eco, the same
cautions apply.  Aymara deserves more serious linguistic attention
and less mysticism.

> <snip: description of the Logos translation system>

> I worked at Logos only intermittently, so whether they finally got
> it straightened out or not I don't know, but another thing was that after I
> left the system degraded terribly, I came back for a special job several
> years later.  

> So this is my history, just
> how it would apply to the possibility of Aymara serving as an interim
> language I'm not sure but certainly the precision of expression Aymara
> allows is a very remarkable feature.  

There have been several attempts to use a natural language as an
intermediate language for machine translation, notably the BSO project,
which used Esperanto, and Guzmán's Atamiri, supposedly based on
Aymara.  In fact, Esperanto has its own quirks and ambiguities, so
the BSO project in fact used a modified form of the language, with
extra bracketing, for example, to show what modified what.  Whether
Guzmán's project really used an Aymara interlingua has been questioned.

> And I know that my
> Logos experience would be very valuable, because we were the first to
> actually do this.  

> There were other people getting into computerized
> language translation, but no one else came up with so much and such complex
> resolution as we did. 

Considering the long and varied history of machine translation (a
field in which I worked for six years) this is highly unlikely.
Logos is a typical second-generation machine translation company
that has had its ups and downs.  I'm not saying that people at
Logos did not do good work, but just that claims about Logos being
first are going to be very difficult to defend.

The Logos webpages are at  http://www.logos-usa.com/  

Do not confuse it with another Logos in Italy www.logos.it

The history of Logos is potentially interesting, and I have been
trying for some time to get more reliable information.
To begin, it would seem that the founders Bud Scott and 
Charlie Byrne, and their first workers, belonged to a "Jesuit
sect" that lived on a "commune" somewhere near Middletown,
New York, USA.  Presumably their primary motivation was
to support missionary work.  I have this information second-hand,
so it needs further checking.  Corrections and additions to this 
account would be welcome.

Another popular machine-translation group, Globalink, _may_ 
have connections to Georgetown University (again Jesuit).
Again, corrections and more information would be welcome.
Very early MT work at Georgetown also led to the Systran MT system.

> so I contacted the name, a Professor
> Hovy, and learned from him that nowadays they let the computer do the
> translating.  They have to start with a source document and a target
> translation of it, which they feed to the computer, which says to itself,
> whenever I see this word "four" in English, for instance, it is "quatre" in
> French.  But of course this isn't really "translation" it's just a
> dictionary lookup, with no grammatical resolution. 

This is almost certainly _not_ what Hovy said or meant.  He is a very
well informed and sophisticated scholar of machine translation.  Older MT
systems/projects like Weidner, Logos, METAL/T1, TAUM-METEO, Eurotra, etc.
tried to do syntactic and semantic analysis and generation using
hand-written linguistic rules.  This approach proved generally "brittle",
producing systems that failed to analyze many real sentences found in
real corpora, and producing rather disappointing output.  (I say this
with some regret, because my own training and inclination is towards
rule-writing.)  The newer approaches, which was what Hovy was probably
describing, are based on "translation by example" and use statistical
techniques.  The rules to be followed when translating from French
to English, for example, are automatically derived from extensive
aligned bilingual French-English corpora.  Right now, in natural
language processing in general, the statisticians have the upper hand
over the more traditional rule-writers (like me).  Advances in web-search
technology and especially speech recognition came about when the
projects went statistical (and got rid of a lot of rule-writing
linguists like me).  Perhaps the pendulum will swing back soon, but
go to any ACL (Association for Computational Linguistics) or
COLING conference and you will see the statisticians everywhere.


Kenneth R. Beesley		ken.beesley@xrce.xerox.com 
Xerox Research Centre Europe    Tel from France:    04  76 61 50 64  	
6, chemin de Maupertuis		Tel from Abroad: +33 4  76 61 50 64
38240 MEYLAN			Fax from France:    04  76 61 50 99
France				Fax from Abroad: +33 4  76 61 50 99

XRCE page:      http://www.xrce.xerox.com
Personal page:  http://www.xrce.xerox.com/people/beesley/beesley.html