Kick off Posting...
shash...@gmail.com
Thanks for starting off this forum and the invite. I am looking
forward to some interesting discussions along applying HTM concepts to
linguistics (natural language understanding?). I am curious as to the
state of research and what ideas you and others have been
contemplating, so let me start off by sharing what I have been
doing....
My work at MIT was primarily on generating Semantic Web Ontologies
from text corpora , for which I wrote an LSA-based system to spit out
a bunch of RDF triples. I am interested in using HTM to build a
"reasoning" with the triples, possibly as a complement to DL/FOL
reasoners. Has anyone been thinking or working along these lines?
Look forward to hearing your ideas..
Regards,
Shashi
James Tauber
Welcome Shashi
Like Jeff Hawkins, I live a double life, although I'm not quite as
successful as him in either life :-)
By day I'm a software entrepreneur: Chief Scientist and co-founder of a
company called mValent. By night (and weekends) I'm a PhD student:
linguistics part-time at University of Essex.
I'm a straight-up linguist rather than a computational linguist or
natural language processing guy. That said, though, being a software guy
by vocation, I do a lot of coding around the linguistic models I deal
with. My doctoral studies are focused on morphology and, in particular,
the inference of inflectional morphology. I'm not a psycholinguist
either, although I'm interested in psychologically-motivate models of
language.
I've wanted to start this mailing list for a long time - since I first
read On Intelligence in 2005. I've only now gotten around to doing it
:-)
The first obviously application of HTM I thought of was in something
like part-of-speech tagging where you are making predictions within
sequences where you have cues from above (syntax) and below
(morphology). There may be obvious applications in speech recognition
too, but that's beyond my expertise.
James
--
James Tauber http://jtauber.com/
journeyman of some http://jtauber.com/blog/
shash...@gmail.com
being able to juggle all these at once.
I see your company is US based and you are in the UK.
Have you actually done any experimentation in using syntax and
morphology to identify parts-of-speech?
That seems a very interesting challenge.
James Tauber
On 07/02/2007, at 2:41 PM, shash...@gmail.com wrote:
> That is a very interesting profile and background. Congratulations on
> being able to juggle all these at once.
> I see your company is US based and you are in the UK.
Lately I've mostly been in the US - spending next week in UK and a
couple of weeks in March in Australia, though.
> Have you actually done any experimentation in using syntax and
> morphology to identify parts-of-speech?
> That seems a very interesting challenge.
The word form itself as an input to POS-tagging is pretty much a non-
starter in a language like English. Which is why I suspect it doesn't
get used much.
Most POS-tagging approaches seem to be just based on n-gram
probabilities, maybe some Hidden Markov stuff (Viterbi, etc)
But, in a language with rich inflectional morphology, if one adds
input from an analysis of the form itself as well the surrounding
context, one might get better results. And this approach is naturally
implemented in an HTM-like manner (i.e. layers of sequences with
conditional probabilities between the layers and the need for a Pearl-
like belief propagation algorithm).
So part-of-speech tagging is the first linguistic application of HTM
I'm going to try. It's not really my area of speciality but it seems
a nice fit and there are plenty of other approaches with objective
measures of success that the HTM approach can be compared to.
James