In wrapping up the section on representing lexical knowledge, issues directly related to word-sense distinctions will be discussed, because this forms basis for the the thesis research. Kilgarriff [65] considers the ``bank model'' of sense distinctions, in which word senses are clearly delineated, to be inadequate. Furthermore, he takes issue with the claim that words only have a single meaning in context. To support his viewpoint, he analyzed a set of LDOCE definitions and characterized the distinctions that were present. The results indicate that a significant portion of the distinctions are not those that can be readily reconciled with the clearly-delineated senses approach. However, the results are only preliminary and only account for 24% of the pairwise distinctions (or 42% of the individual sense distinctions).
In an additional study, usages from a corpus were matched with the closest related sense for a set of 83 words. For the words in which part-of-speech doesn't resolve ambiguities, he found that 87% words didn't have usages which could clearly be resolved. Note that a strict test for ``clearness'' is used since all citations for a word must be resolved.
Hirst [59] addresses subtle word-sense distinctions to account for near synonyms (called plesionyms). The approach sketched out is to represent the differences among the plesionyms as objects in order that they can be manipulated directly.
He suggests a two-level knowledge representation scheme, modeled after proposals common in the literature. Course-grained conceptual knowledge will be stored in a taxonomy, and fine-grained language-specific knowledge is stored in the lexicon, which is similar to the scheme used in Mikrokosmos [90]. For plesionyms that represent distinct concepts, differences can be determined by comparing attributes, including those inherited from ancestors leading to a common ancestor.