Re: SUO: Re: Enhancing Data Interoperability with Ontologies...
> John, how does that compare to human subjects
> reading sentence sequences and marking anaphora
> manually? Do humans do very much better, or
> about the same?
Slight correction: Mitkov is at Woverhamption not
Southamption as I said in an earlier email: sorry,
those -hamptons all sound the same to me! :-)
The following:
http://www.clg.wlv.ac.uk/papers/mitkov-98b.pdf
is a paper reporting 89.7% fully automatic accuracy for technical
manuals adopting a deliberately "knowledge-poor" approach.
It contains references to some of the standard approaches
in the area.
There have been frequent workshops/conferences on the
task:
e.g.:
Operational Factors in Practical, Robust, Anaphora Resolution for Unrestricted Texts
ACL'97/EACL'97 Workshop on Intelligent Scalable Text Summarization Madrid, Spain, July 11 or 12, 1997
I still haven't found clear tables for results of
human performance: so if anyone comes across some,
please let me know too! In general, though, the
figure is going to be very high for the forementioned
reasons. There will be "errors" only where there is
complete ambiguity or less than fully accurate
clues given in the text and there is a lack
of necessary disambiguating knowledge on the
part of the hearer/reader. There may well be
other figures that can be ascertained by doing
experiments under varying cognitive loads (i.e.,
understanding complex anaphors while driving
a racing car and the like): success will also
go down here but how far and how quickly? no
idea. Somebody must have studied it somewhere:
not my area. So I won't opine (any more than
I have :-).
John B.
--
John Bateman
FB10, Sprach- und Literaturwissenschaften
Universität Bremen
28334 Bremen, Germany.
Tel: +49/421-218-9483
Fax: +49/421-218-4283 (or 218-7801)
http://www.uni-bremen.de/~bateman