W3C home > Mailing lists > Public > semantic-web@w3.org > October 2005

Re: sentence in NLP to ontology query?

From: Joshua Tauberer <tauberer@for.net>
Date: Mon, 24 Oct 2005 09:31:37 -0400
Message-ID: <435CE239.5020808@for.net>
To: Xavier Noria <fxn@hashref.com>
CC: Semantic Web <semantic-web@w3.org>

Xavier Noria wrote:
 > I have seen in several places those questions in natural language that
 > are said to be addressable with, say, the expressiveness and logic
 > properties of OWL DL.

Good question. :)

The first problem in running a query given a NL sentence is figuring out 
what the NL sentence is asking -- coming up with some representation for 
the meaning of the sentence.  There are maybe two broad ways you could 
look at the problem.  One is sort of the engineering perspective: Find 
any algorithm that works for the questions you expect to encounter.  The 
more limited your domain, the easier that is to do.  That's my sense of 
the approaches taken in the citations Philipp posted (not having read 
any of the papers).

The second view is the linguistic one: Find an algorithm like what we've 
got in our brains that will compute the meaning of any sentence.

Going the linguistic route is, well, difficult, to say the least. 
Linguists studying semantics are trying to solve this problem.  And, the 
punch line is that no one really knows.  There are lots of good ideas, 
but there's no complete solution to figuring out what a sentence means.

I'll give an example of one of the issues using the example you 
initially gave:

> Tell me what wines I should buy to serve with each course
> of the following menu. And, by the way, I don't like Sauternes.

The first sentence has two readings (and more I'm sure, but here are two):
   1) What set of wines W satisfy the following: I should serve all of 
the wines in W with each course.
   2) For each course, tell me which set of wines W should be served 
with that course.
That is, 1 is asking for one set of wines compatible with all of the 
courses.  2 is asking for a set of wines separately for each course.

Now, the problem of ambiguity is huge in itself, but what I want to 
illustrate is that whatever algorithm one comes up with, it has to 
recognize that these two readings are separate and both possible.  (This 
is an issue of 'quantifier scope'.)

Above the level of semantics there's the issue of pragmatics.  A good 
example of that is the second sentence in your example.  The literal 
meaning of "I don't like Sauternes" is irrelevant to the query.  The 
user was implicating something else, something like "And exclude from 
the result all Sauternes."  How to go from the literal meaning to the 
implicature is another story entirely.

Below the level of semantics there's the issue of syntax, and the moral 
of the story here is that there's no agreement on even what the 
structure of a NL sentence is.  (This happens to be what I'm studying.)

So, this is why there's aren't any algorithms out there for transforming 
NL sentences like the one you posted into queries -- at least, not from 
the linguistic perspective.  It's pretty tough.  The domain of possible 
questions has to be significantly restricted before one can make any 
headway.  I'm optimistic that some good progress will be made in 
linguistics to get closer to solving this, but realistically it's far off.

-- 
- Joshua Tauberer

http://taubz.for.net

** Nothing Unreal Exists **
Received on Monday, 24 October 2005 13:31:52 UTC

This archive was generated by hypermail 2.4.0 : Friday, 17 January 2020 19:47:07 UTC