Re: Question on Stochastic Language Models (N-Gram) Specification WD 3 January 2001

Paul van Mulbregt wrote:
> 
> Thanks for your long and thoughtful reply, James, but I had actually meant
> the question in a much more nuts-and-bolts way. (My own background is LM
> R&D at the former Dragon Systems).  Rather than immediately list reasons
> why I didn't think the spec solved what I thought it was trying to solve, I
> wanted to find out what the authors thought the problem was.  And I wanted
> to follow the data all the way to the platform seeing exactly what the
> authors envisioned happening at each step.
> Cheers,
> Paul
> -------------------------------------------------------------
> Paul van Mulbregt,  paulvm@ne.mediaone.net

BTW, I probably should also mention IBM has n-gram technology.  I don't
know what their plans are for using it.

Getting into the platform architecture, a web-based system using n-grams
will need to define a semantic interpreter right now.  We have proposed
a simple tagging mechanism for n-grams, but the n-gram model poses a
more difficult problem since higher-level structure is statistically
derived.

The motivation for this spec comes from requirements established early
in the process at the request of several members of the working group. 
The main target was support for dictation style tasks, but as I
mentioned in the last email, there are many non-dictation applications
that still want unrestricted NL interfaces.  This spec allows the LM to
be speced on the web server with other application information.  We
still have work to do on bringing in the semantic spec, so this is in
the future study section right now.  My guess is we will work more of
this out post-VoiceXML 2.0 after we see how the semantic markup turns
out for the BNF format.  Of course, you are welcome to make suggestions.

	Mike
-- 
		Michael K. Brown
		Avaya Labs, Rm. 2D-534, (908) 582-5044
		600 Mountain Ave., Murray Hill, NJ 07974
		mkb@avaya.com

Received on Wednesday, 14 February 2001 11:34:03 UTC