Lyons

Lyons lyons:77 devotes a section to the discussion of the homonymy-polysemy distinction (p. 550-569). He identifies the following criteria as those traditionally applied in making the distinction:

1. etymological information -- homonymous lexemes ``should be known to have developed from what were formally distinct lexemes in some earlier stage of the language''
2. relatedness of meaning -- homonymous lexemes have unconnected meanings

Lyons correctly criticises the first criterion as being irrelevant to the synchronic analysis of language, since native speakers are largely unaware of the etymology of the words of their language yet they are able to assign meanings to them. Notice that this criterion will not exclude bank as discussed above from being polysemous, given that the two senses historically derive from a single lexeme and in fact a single sense. The second he identifies as an important consideration, but points out that relatedness of meaning is a subjective measure for which intuitions may vary among individuals.

Lyons considers two alternatives to circumventing the homonymy-polysemy issue:

1. Maximise homonymy -- associate every meaning of a word with a distinct lexeme. Lyons shows that this will lead to considerable redundancy in the lexicon, as much morphological, syntactic, and even semantic information will be repeated in the lexical entries for the distinct lexemes. However, this redundancy can be greatly reduced given current inheritance-based approaches to lexicon construction. More problematic is the observation that this approach depends on the ability to spell out in advance all of the possible senses in which a word will be used. Lyons suggests that sense distinctions can be ``multiplied indefinitely'' (1977:554) and that therefore this tack is hopeless. It will never be possible to decide in advance the full range of possible senses a particular word might be associated with, and furthermore it makes the computational task of selecting the appropriate lexeme daunting given the number of lexemes which might be associated with a particular word form (see Section 6.4 below).
2. Maximise polysemy -- adopt the notion that no two lexemes can be entirely distinct when they are syntactically equivalent and when the set of word forms they are associated with are identical. On this view, there are only various kinds of partial homonymy (i.e. when there exist syntactic differences among uses of a word). This removes the vague concept of ``semantic relatedness'' from the lexicon. However, it would result in an extremely underspecified lexicon from which very little information about the meaning of words could be gleaned. It suffers from the problem of an inability to explain the intuitions that underly the notion of homonymy, and, more relevant to computation, from a complete inability to identify the normal context of use of a particular word and no basis for establishing synonym classes or other semantically-based groupings. How any useful interpretation could be accomplished without some sense differentiation is difficult to see.