UMIACS Computational Linguistics Colloquium Series,
December 11, 1997
UMIACS Computational Linguistics Colloquium Series,
December 11, 1997
Stochastic Attribute-Value Grammars
Steven Abney
AT&T Research
Stochastic grammars are a topic of central interest in current
computational linguistic research. Stochastic regular and
context-free grammars are widespread, and parsing and
parameter-estimation methods for them are well-known. There have been
attempts to extend these methods to attribute-value grammars, but none
of them provide a correct parameter-estimation method, in the sense of
one that converges to even a local maximum of likelihood.
In this talk, I show that the Improved Iterative Scaling procedure of
Della Pietra, Della Pietra, and Lafferty can be adapted to provide a
correct parameter-estimation method for stochastic attribute-value
grammars.
Return to the UMD
Computational Linguistics Colloquium Series.