The conclusion for the computational lexicographer which can be drawn from the previous discussion is that the homonymy-polysemy distinction may not in fact be a useful one for the purpose of designing a lexicon to be used in an NLP task. It does not suffice as a basis for decisions about which words require several lexical entries, and which words can simply be represented by a single, underspecified, lexical entry, because the kinds of variations in meaning associated with a word do not fall neatly into two groups which might be called ``variations which can't be explained semantically'' (homonymy) and ``variations which seem to have some semantic basis'' (polysemy). Given that most NLP systems that require some interpretation capability will have need for making some distinctions between different meanings associated with a word, how might we go about representing these distinctions once they have been decided? The grain of distinctions will, however, have to be chosen with regard to the application of the system, as I suggested in Section 6.2.