Croft's account of word classes is a combination of semantics and pragmatics. Briefly, he suggests that crosslinguistically valid categorisations of nouns, verbs, and adjectives should be seen as the three ontological categories cross-cutting with Searle's (1969) three major propositional act functions. Specifically, nouns are object words that perform the function of reference, verbs are action words that predicate, and adjectives are property words in a modifying function.
Objects, actions, and properties are further analysed into the dimensions of relationality, stativity, transitoriness, and gradability, see e.g. Croft (2001: 87). As for the propositional acts, reference "identifies a referent and establishes [or maintains, WBH] a cognitive file for that referent, thereby allowing for future referring expressions coreferential with the first referring expression"; predication "ascribes something to the referent" and modification "(of referents) functions to enrich a referent's identity by an additional feature of the referent, denoted by the modifier" (Croft 2001: 66).
In addition to the noun, verb, and adjective prototypes, Croft also discusses non-prototypical combinations. For example, Palmer's (1971) example of his suffering would be seen as an action word used as reference. Note however, that compared to prototypical nouns it has extra morphological marking (gerundial -ing), and that it cannot take a plural inflection. Croft argues that this is not a coincidence precisely because this combination is non-prototypical; for full details concerning the typological generalisation underlying these observations see (2001: 90-92).
Croft draws some interesting parallels between Langacker's summary and sequential scanning on the one hand, and reference and predication on the other. In his own words:
Langacker's concept of sequential scanning for predications (...) is in some ways the opposite of opening [or maintaining, WBH] a cognitive file. In opening [or maintaining, WBH] a cognitive file, the entity is activated as a whole (Langacker's "summary scanning", the opposite of sequential scanning) and remains active, at least for a while. In sequential scanning, however, the temporal phases of the entity are focused on and then dropped as soon as the new phase (and also a new event) is scanned. (Croft 1991: 121)
The same point is made by Croft & Cruse (2004: 54).
If we assume that Langacker's (2008b) revision of his account of scanning (see Section 2.2.1, above) is on the right track, a more subtle conception of Searle's propositional act functions is called for as well. Specifically, predication is not simply an all-or-none affair, but it may be more or less salient, depending on the independence of the clause. Croft himself actually provides a hint in this direction, when he states that "[fjurther research on the grammar of (...) nominalised (or at least deverbalized) verb constructions is necessary" (1991: 121). He actually also discusses noun phrases whose function is not referential, e.g. nouns that are incorporated into verbs (Croft 1991: 121), as well as predicate nominals, "which are commonly inflected just as verbs are in a language" (ibid.:120).
Croft's definitions of the propositional act functions given above make no reference to their (morpho-)syntax. In fact, he explicitly denies that these functions are to be seen in structural terms: "these grammatical functions are definitely not formal. Instead, it is best said that they are pragmatic functions" (1990: 248). For purposes of crosslinguistic comparison this position is inevitable, but in a particular language the functions will nonetheless often be represented by certain constructions. For example, Croft suggests that "man is a prototypical noun only when it is the head of a referring expression such as the man; ran is a prototypical verb when it is the main predication, as in the man ran" (1990: 247-248). These constructions often constitute criteria for nounhood and verbhood in generative accounts as well; cf. Section 2.1, above.
In a usage-based view of language, to which Croft subscribes as well, the frequent occurrence of nouns and verbs in such constructions would come to be stored as part of speakers' knowledge of these categories. In fact, regardless of the fact that these constructions are language-particular, Radical Construction Grammar would need to propose that in individual speakers' grammars these constructions provide the very basis for the word classes. In other words, it would seem that the Radical Construction Grammar account of word classes overlaps to some degree with the traditional generative view. Constructions that are clearly associated with the propositional acts of reference and predications will form the basis of the distributional analysis carried out in this study. This is discussed in Section 3, below, with Section 5 reporting on the results. Propositional acts will be treated there also in light of Langacker's revised account of summary and sequential scanning (see Section 2.2.1, above).
Psycholinguistic research on lexical categorisation has confirmed the importance of both distributional and semantic criteria. Early evidence for distribution was supplied by Berko's well-known (1958) wug study. Regarding semantics, Brown (1957) showed that subjects associated nonsense words such as sib with either objects or actions depending on the context they appeared in: compare a sib to sibbing. This study therefore actually also confirmed the role of distribution.
Some work on distributional criteria does so without any regard for constituency or meaning. In other words, the syntactic frames are identified purely on the basis of frequency, regardless of whether the pattern is particularly meaningful or complete. The frames in question are often bigrams (e.g. Monaghan et al. 2005) but sometimes slightly larger units (see e.g. Mintz 2003, who uses X_Y frames). This results in the inclusion of patterns such as a with (Monaghan et al. 2005: 162) or you_a and why_you (Mintz 2003: 102), which are not constituents in a traditional linguistic sense and do not dovetail well at all with the criteria proposed by generative grammarians, which I have hinted in Section 2.2.1 may be compatible with cognitive linguistic research.
Not all psycholinguists take this 'blind' approach. Höhle et al. (2004), for instance, looked at whether the acquisition of nouns and verbs in German children aged between 12 and 16 months was supported by the presence of determiners and subject pronouns. They find that at 14-16 months the presence of determiners helps children identify nouns, while subject pronouns do not (yet) play a role. They explain this by adducing corpus evidence that shows that in German, with its relatively free word order, subject pronouns have a much lower cue validity for verbs than determiners have for nouns (Höhle et al. 2004: 349).
Regardless of the exact operationalisation of distributional cues, it is striking that psycholinguistic research has been largely ignored by theoretical linguists, generative and cognitive alike.
In addition to semantic and distributional criteria, since the late 1980s psycholinguistics has yielded a now substantial body of evidence for the role of phonological cues in lexical categorisation (e.g. Cassidy & Kelly 1991, 2001, Don & Erkelens 2008, Durieux & Gillis 2001, Farmer et al. 2006, Kelly 1992, 1996, Kelly & Bock 1988, Monaghan et al. 2005, Sereno & Jongman 1990, Shi et al. 1998). Just as in the case of the psycholinguistic literature on distributional criteria, researchers have tackled the questions as to whether cues are in principle available, whether they can be used in categorisation, and whether they are indeed used. Corpus data, artificial learning algorithms, and data from real learners/speakers suggest that these three respective questions should all be answered in the affirmative. The cues in question (for English) are discussed in Section 3 below.
Once again, it is surprising that theoretical linguists have by and large neglected to incorporate this phonological evidence into their models. Two notable exceptions in the cognitive camp are Berg (2000) and Taylor (2002). In fact, Taylor's suggestion that in between specific nouns and verbs and the maximally abstract noun and verb schemas there are phonologically partially specific "sub-schemas" (2002: 184) anticipates the proposal made below.