The traditional view of lexical organisation, evident both in lexicography and linguistics (e.g. Kempson 1977:79-83), is that distinct word senses should correspond to distinct entries in the lexicon. The arguments in favor of this position hinge on the discreteness of the meaning expressed in a usage of a word -- where there is ambiguity, only one sense of a word can be active at any one time. This perspective explicitly denies that the representation of the meaning of a word can be underspecified and modulated (to borrow Cruse's terminology) or specified in context.
Under this view, polysemy is indistinct from homonymy, and so it exemplifies Lyons' ``Maximised homonymy'' (Lyons 1977; see previous section). It is therefore subject to the criticisms Lyons puts forward: it leads to much lexical redundancy, and it demands the impossible task of enumerating in advance all the senses which might be associated with a lexical form.
Given the productivity and dynamism of language and word meaning, this approach is simply inadequate as a lexical semantic theory, and therefore inadequate for any complex computational language interpretation. It treats the lexicon as independent from grammatical or pragmatic processing; the lexicon is simply a static repository of information which is accessed in, but cannot be influenced by, such processing. It restricts the effect of context on the lexicon to sense selection. Consider example cc1 above, repeated and expanded here as cc5, in which different parts of a car can be picked out by different verbs. To handle this on the multiple lexical entry approach, there either need be distinct lexical entries for car corresponding to its different parts, which then wouldn't account for a single instance of the word which encompasses two of these lexical entries, as in cc6, or there could be a single general lexical entry for car which was neutral to which part of the car was being referred to, but this would mean that a lot of information about the interpretation of these sentences would be lost at the lexical level. The latter solution would mean that the specific interpretations would have to be left to (computationally expensive and) complex pragmatic reasoning.
The car needs servicing. (the engine) The car needs washing. (the body) The car needs rust-proofing. (the under parts)
The car needs washing and servicing.
Furthermore, to account for examples of metonymy, or transfers of reference (Nunberg 1978), for which a commonly cited example is shown in cc7,
The ham sandwich is at table 7.
ham sandwich would have to have a lexically specified interpretation as ``person who ordered a ham sandwich''. Given the fact that reference transfer is an entirely context-dependent phenomenon, it is difficult to see how it could plausibly be accounted for in terms of a fixed lexicon.
From the perspective of computation, the multiple lexical entry model is quite attractive in that it assumes a finite set of choices from which the NLP system can choose the most appropriate. It is as a result the most widely used model for the computational lexicon to date, and word sense disambiguation research has been framed entirely within this model. This approach will likely suffice for highly-specific domains, in particular when the relevant senses are well-defined with respect to one another, and NLP tasks which do not need to approach human competence in sense discrimination. In fact, highly domain-specific tasks might suffer under the alternative approach of a fully generative lexicon due to the loss of a domain-specific information. However, this could be addressed by adding domain-relative word sense frequencies to the general generative lexicon.