Abstract: This paper describes automatic techniques for mapping 9611
semantically classified English verbs to WordNet senses.  The verbs
were initially grouped into 491 semantic classes based on syntactic
categories; they were then mapped into WordNet senses according to
three pieces of information: (1) prior probability of WordNet senses;
(2) semantic similarity of WordNet senses for verbs within the same
category; and (3) probabilistic correlations between WordNet
relationship and verb frame data.  Our techniques make use of a
training set of 1791 disambiguated entries, representing 1442 verbs
occurring in 167 of the categories.  The best results achieved .58
recall and .72 precision, versus a lower bound of .38 recall and .62
precision for assigning the most frequently occurring WordNet sense,
and an upper bound of .75 recall and .87 precision for human judgment.
Acknowledgements: All three authors are supported, in part, by
PFF/PECASE Award IRI-9629108, DOD Contract MDA904-96-C-1250, and DARPA
Contracts N66001-97-C-8540 and N66001-00-2-8910.  Rebecca Green is
supported, in part, by a National Science Foundation Graduate Research
Fellowship.  We are indebted to Philip Resnik for his assistance with
experimental runs of his algorithm on the data and his useful
commentary in the preparation of this document.
(Submitted to 39th Annual Meeting of the Association for Computational
Linguistics, Toulouse France.)

(Also cross-referenced as LAMP TR-061)
(Also cross-referenced as UMIACS TR-2001-02)