There are some other considerations in these experiments ... Reuters is GP text, so more ambiguities in words whereas CACM is fairly domainspecific and thus wsd is easier or words will always have only 1 sense in such restricted domains ... some lexicon vocabularies have finer-grained senses than others, eg WordNet is notorious for this.
(much of the following discussion is abbreviated from recent postings to
the CORPORA mailing list)