The cognitive linguistic approach
In the area of word classes the two main cognitive linguistic theorists are Langacker and Croft. Their positions will be discussed in turn.
Langacker's definition of word classes is usually represented as being entirely based on semantics (see e.g. Kelly 1992: 350, Taylor 2002: 179-180, Evans & Green 2006: 555, Verhagen 2009: 126). This is almost but not completely correct.
In terms of nouns and verbs, the focus of the present chapter, Langacker distinguishes between THINGS and PROCESSES. The capitals suggest that these notions are not to be equated with the simplistic descriptions given in traditional grammars such as Nesfield (1908), see Section 2.1, above. Thus, rather than "some individual person or thing or some kind of person or thing", a THING represents "a region in some domain" (Langacker 1987a: 189), or in other words "any product of grouping and reification" (2008a: 105). A PROCESS, by contrast, is a relation between entities which develops and is mentally followed through time (see Langacker 1987a: 248). The latter, but not the former, involves so-called
"sequential scanning", whereby the configurations of the relation are tracked from one moment to the next. In Cognitive Grammar, sequential scanning is opposed to "summary scanning", in which the various configurations of a scene are made available as a single Gestalt. Langacker often likens sequential scanning to a film of a ball flying through the air, with a picture that represents its complete trajectory as an arc corresponding to summary scanning.
Recently, in response to criticism from Broccias & Hollmann (2007), Langacker (2008b) has modified his descriptions of the two scanning modes. The distinction between the two scanning modes is no longer seen as necessarily sharp. More precisely, the degree to which sequential scanning is evoked by a verb depends on whether it heads a finite clause or is somehow subordinated, e.g. when modifying a noun or when functioning as a non-finite complement. In such cases, sequentially may be partly or wholly suppressed, leaving behind a summated view of the event (Langacker 2008b: 576). The exact relation between subordination and different degrees of attenuation of sequentiality is still not made fully explicit, but I attempt to implement this suggestion in interpreting the distributional data; cf. Sections 3 and 5, below.
One respect in which Langacker's account is not necessarily purely semantic is that he tends to hedge his claims by suggesting that nouns and verbs are in principle "semantically definable" (2008a: 108; see also e.g. 2002: 100). A comprehensive alternative characterisation is not offered in Cognitive Grammar, but Langacker has made occasional suggestions about phonology and distribution. Especially his remarks on the latter may be interpreted such that he is not oblivious to its importance.
In terms of phonology, Langacker argues that the realisation of the noun and verb schemas are constrained firstly by the limitations of "phonological space" that is "our range of phonic potential, i.e. our capacity to deal with sounds, and with speech sounds as a special case" (Langacker 1987a: 76). Thus, for example, the noise of a dog barking could never be part of the phonology of a word in human language. The second phonological constraint consists in the language-particular phoneme/allophone and phonotactic schemas that emerge as generalisations over language use (see e.g. 2002: 272-277, 291). For instance, speakers of (present day) English learn that words do not start with /kn/, so that knight is pronounced [naitj. In other words, the noun and verb super-schemas must be linked to (universal) phonological space, and they must also interface with the phoneme, allophone, and phonotactic pattern schemas of a given speaker's language. Note, however, that this does not help to differentiate between nouns and verbs; it merely sets possible words in a language apart from impossible ones.
With respect to distribution, words are defined in terms of the constructions they occur in (Langacker 2005, 2008a: 97, 240-244). The more different frames the item occurs in, the higher the degree of autonomy it will obtain in a process called "decontextualisation". Thus, the noun kith has very little autonomy, as it is restricted to the idiom kith and kin. A noun like house, by contrast, is far more autonomous. It occurs in certain idioms, such as a house of cards and get on like a house on fire, but it can be used in many other contexts too. Be that as it may, Langacker suggests that decontextualisation does not result in the meronymic links with the larger constructions being severed (2008a: 241), so house is still connected to constructions such as house of cards.
Although the implications of this position regarding noun and verb schemas are not explicitly discussed, the usage-based nature of Cognitive Grammar is not incompatible with the suggestion that nouns and verbs are connected to the constructions they typically occur in which, to anticipate the discussion in Section 4, may well be the very ones identified by generative grammarians. Thus in relation to word classes there may be more common ground between Cognitive Grammar and generative grammar than is often thought, although the latter does of course typically ignore the role of semantics (at least after early acquisition; see Section 2.2.1, above).
-  This is a slight oversimplification of summary scanning: prior to the Gestalt being available there is a short, incremental "build-up phase", during which the summated scene is constructed (1987a: 251).