From: David Picon Alvarez (eleuteri@myrealbox.com)
Date: Sun Dec 04 2005 - 01:38:36 MST
From: "Herb Martin" <HerbM@LearnQuick.Com>
> Do you have any particular references (preferably online)
> for "text summarization" -- that is, the ability to extract
> summaries or to paraphrase text and information?
Not exactly text summarization, but I'd advice you to check the Pascal's
(nothing to do with the language) automated recognition of textual
entailment challenge.
http://www.pascal-network.org/Challenges/RTE/Introduction/
The idea is to generate an automatic process by which a sentence can be
recognized to be logically entailed from another. For example: Jane works as
an electrician in the weekends entails Jane is an electrician, Jane is a
human, etc. (Note that entailment itself generally has some common sense
issues, Jane could be a robot, but we'd hardly consider this at the moment.)
I was doing some undergraduate-level work on this stuff, and the existing
approaches are very shallow, semantically speaking. The common thing to do
is using very very low-level metrics like Lvenshtein distance and such in
order to find markers of entailment, and the only group I know of (at
Edimburgh University) which went about it the Right Way (tm), building a
parser and trying to do some serious semantic work, gets a much worse
f-measure than the statistics-driven groups. Of course they get very very
good accuracy, but very bad recall, since their semantic analysis isn't
complete and many true entailments pass them by, but when they see one it is
almost sure there is one.
> (This is a field that interests me greatly. Along with
> the idea of programs that write and optimize programs
> seems to be where much of the leverage can be found; and
> where my own VERY modest efforts are directed.)
--David.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:54 MDT