Cruse (1986) distinguishes lexemes from lexical units. The former are ``the items listed in the lexicon, or `ideal dictionary' of a language.'' A lexeme corresponds to a particular word or word form, and can be associated with indefinitely many senses. The latter are ``form-meaning complexes with (relatively) stable and discrete semantic properties'' (p. 49), and the meaning component is called a sense, corresponding to the intuitive notion of sense I have been using. So bank is a lexeme, while bank-financial institution and bank-edge of a river are lexical units.
In discussing the semantic contribution of a word to a sentence, Cruse differentiates between contextual selection of a sense and contextual modulation of a sense. Selection refers to activation of a particular sense of an ambiguous word form due to the context, while modulation refers to variation induced by the context in terms of emphasising or de-emphasising various aspects of the sense. The sentences in cc1 exemplify modulation in that the two sentences highlight different parts of the car (the engine and the body, respectively) rather than requiring that car refer to different entities in each case. The sentences in cc2, on the other hand, are instances of contextual selection: a different sense of light is selected in each case.
The car needs servicing. The car needs washing.
The room was painted in light colours. (cf. dark/*heavy) Arthur has rather a light teaching load. (cf. heavy/*dark)
Contextual selection corresponds to the kind of word sense disambiguation undertaken in most NLP systems -- there is a pre-existing set of senses for a lexeme (word) and the relevant sense must be identified based on the context. Contextual modulation refers to one creative aspect of language use, and points to the need for complex representation of knowledge about a word. This is because multiple aspects of a word can be activated simultaneously: The car needs servicing and washing is completely felicitous despite the fact that different facets of the car are referred to by each of the verbs (cf. ??Arthur has light teaching loads and rooms in his house which indicates that multiple senses of light cannot be active at once). This kind of knowledge is directly relevant for discourse processing (e.g. anaphor resolution, for example in a discourse such as ``The car needs servicing. It also needs washing.'' where the anaphor refers back to the car as a whole, not just the engine) and generation of coherent sentences/discourses.
Cruse (1986:68) introduces the concept of a gradient of establishment of senses. By this he means that a lexical form can be associated with some senses which are potential rather than explicitly represented in the lexicon. Context can stimulate rules which generate an appropriate sense. This idea serves as the foundation of recent work on the Generative Lexicon (Pustejovsky 1991, pustejovsky:95a: see below). Cruse furthermore argues for the existence of sense-spectra, in which the senses of a lexical form lie along a continuum, with no clear boundaries between them, and in some cases without an encompassing sense. copestake_briscoe:95 provide an example, shown in cc3.
That book is full of metaphorical language. That book is full of long sentences. That book is full of spelling mistakes. That book is full of typographic errors. That book has an unreadable font. That book has lots of smudged type. That book is covered with coffee.
As they point out, co-predication of the first and the last properties seems odd cc4a while co-predication of adjacent pairs seems natural, e.g. cc4b.
?That book is full of metaphorical language and is covered with coffee, so it's very hard to read. That book is full of typographical errors and has an unreadable font.
Cruse (1986:73) suggests that the description of sense-spectra is problematic since a full sense-spectrum does not function as a single lexical unit. Yet he proposes to treat them as a lexical unit, with recognition of the senses along the continuum as local senses.
A lexeme, corresponding to a lexical entry, is proposed by Cruse (1986:76) to be a family of lexical units. This family can either correspond to a sense-spectrum, or to a set of senses which can be related to one another via regular lexical semantic relationships (captured by lexical rules). Thus the structure of the lexicon on Cruse's view essentially reflects only productive relationships and groups of senses capturing different aspects of a single entity which cannot be consistently delineated. In proposing this, Cruse focuses on the lexical unit as the primary semantic unit and on the distinction between lexical units, de-emphasising the importance of the word. As such, he skirts the homonymy-polysemy distinction issue, which involves the relationship between lexemes and lexical units. Although the difference between contextual selection and contextual modulation points to phenomena affected by the distinction, he does not propose clear criteria for establishing the distinction. So the computational lexicographer is left with yet another reason for making the distinction but still no basis for making it.