Rob, (01)
I posted an updated version of my slides, which clarify and elaborate
many of the issues. I'll cite them in responding to your note: (02)
http://www.jfsowa.com/talks/ontofound.pdf (03)
JFS
>> The distinction I was trying to make is between a symbolic knowledge
>> representation, which can be translated to and from a natural language... (04)
RF
> Had anyone ever found a way to do that using logic, your argument
> might hold water. (05)
I have criticized many aspects of formal semantics, but on this point
they have succeeded. Their batting average on arbitrary NL texts is
very low, but when they are successful, they can hit a home run. (06)
No statistical software can translate its internal forms to a coherent
phrase or sentence in any NL. But logic-based systems can and do. (07)
RF
> I personally worked for years on an MT system which attempted this.
> We did not succeed. No-one has succeeded since. (08)
That depends on your definition of success. If you mean FAHQMT
(Fully Automated High Quality Machine Translation) on a broad range
of texts, no MT system of any kind has reached that goal. (09)
But for narrow domains, such as weather reports, the METEO system
from the 1980s did a good job of translating English or French source
to a logical form and generating a report in the other language.
But METEO never attempted to translate the more complex commentary
that might accompany a report about unusual or extreme weather. (010)
My recommendation is to develop flexible, adaptable systems that
can handle an open-ended number of domains of any size. And we
have been doing that at VivoMind -- see Slides 75 to 105. (011)
RF
> Humans may be able to get back and forth to a useful degree, but it is
> exactly how they do that which is at issue. (012)
I agree. That is why I went into quite a bit of detail about how people
process language (psycholinguistics & neuroscience) in Slides 13 to 60. (013)
See slides 24-27 about Sydney Lamb's "neurocognitive networks". See his
course notes for more detail: http://www.owlnet.rice.edu/~ling411 (014)
Slides 28 to 31 cover Peirce's logic and an article by the psychologist
Johnson-Laird: http://mentalmodels.princeton.edu/papers/2002peirce.pdf (015)
Slides 42 to 44 discuss background knowledge and mental maps, images,
and models. Those are issues that depend on mappings between images
and the words, phrases, and sentences of NL. People do that very well.
Even current logic-based systems sometimes do that, but not with the
consistency, efficiency, and broad coverage we would like. (016)
RF
> But no one logic can in itself express all the distinctions necessary
> to make the choice. (017)
I agree. For over 20 years, I have published and discussed the need
for an open-ended variety of logics and ontologies. That is the point
of my discussion of Wittgenstein and my recommendations for a hierarchy
of multiple theories to implement something that resembles W's open-
ended language games. See Slides 45 to 74. (018)
RF
> I agree little has been done historically to implement compositionality
> in particular using distributed representation. That doesn't mean
> it can't be done. (019)
I wasn't criticizing the lack of implementations. I was just pointing
out the obvious: Every "bag o' words" method throws away the structure
of sentences, paragraphs, and discourse. There is no way to recover
the missing information by composing multiple vectors that all suffer
from the same lack of structural information. (020)
But you can do very accurate work with vectors that represent
*both* structure *and* ontology. See slides 76 to 80. (021)
I also want to emphasize that I am strongly in favor of using
statistical methods when they are useful. I was just making the
point that, by themselves, they are not adequate to do language
understanding. I say that in Slide 12. (022)
John (023)
_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J (024)
|