2. A model of variability

I argued for a model of grammar (G) which produces a Pool of Variants (PoV), where each variant is a distinct feature complex, with the same semantic interpretation, and with potentially different phonological forms.

1. (1) G→{v₁, … v_i … v_n} (=PoV)

I also assumed a distinction between knowledge of language and use of language (Chomsky 1965), explicitly stating that I took the grammar G to be embedded in a performance model (see, especially, footnote 10, p. 527). One can conceive of the systems of use U as a choice function on the pool of variants, given a context of utterance C:

1. (2) U(PoV,C)=v_i∊PoV

The function U is extremely complex, and is sensitive to all sorts of properties of the elements of PoV: their phonology, their sociolinguistic connotations, whether they have been encountered recently, their frequency of occurrence in the life of the language user who is speaking, whether the language user likes that particular word, etc. It is also sensitive to many aspects of the context of utterance: the information structure of the discourse, pragmatic expectations about the interlocutor's knowledge, social expectations about appropriateness etc.

One further factor, which I made use of in Adger (Reference Adger2006), is that the phonological form of variants might not be evenly distributed across PoV, so that more than one v_i could have the same pronunciation. In this way, the structure of PoV itself has an effect on the phonological form of any particular utterance. I make this clear in my paper in the passage Hudson quotes:

The probability of any particular v_i being chosen in any speech event is a function of the factors which enter into the specification of U and C. The fact that some phonological form might be more common in a corpus than some other phonological form depends on both the factors specifying U and C and the structure of PoV itself. So we must distinguish between the probability of any variant, and the frequency of the particular phonological forms (note how this is subtly different from the classical Labovian notion of linguistic variable). The probabilities of the variants can be given as is standard in probability theory.

1. (3) p(v₁)+ … +p(v_i)+ … +p(v_n)=1

Each v_i here is a pair of syntactic features and phonological form. The probability of occurrence of any particular v_i can be empirically modeled in the way familiar from variationist sociolinguistics (e.g. Labov Reference Labov1972). One could attach probabilities directly to the variants (in an objective probability model), although my hunch is that a Bayesian subjective probability approach is a more useful conceptual framework allowing dynamic updating of U.

The claim I made in Adger (Reference Adger2006) was that, in addition to (3), if there are more variants with one form (say was) than another (say were), then that will further impact on the frequency of was vs. were in a corpus. I then showed how this seems to be correct when one takes a large enough sample from a single speech community (see next section for further comment).

In this model G is Socio-free and Use-free. There is never any rule of grammar that makes reference to frequency of a variant, or to social status of a variant. Grammar is sensitive only to syntactic features and structures built up from them. U, on the other hand, is neither Socio-nor Use-free; however, U does not construct syntactic representations or constrain dependency relations between constituents. Moreover, this general model sees U as a dynamically changing function, responsive to the particularities of the utterance situation, and taking into account all of the multifarious factors that influence the particular choice of variant and that are the subject matter of much sociolinguistic research (hence the point about subjective probabilities immediately above).

One might not like this model, but there is no doubt that it contains a ‘purely linguistic’ module and that it is compatible with the fact of sociolinguistically conditioned variation. I think that this is a straightforward falsification of Hudson's basic claim.

A secondary issue is whether this is a plausible model, given our current knowledge of how language works and how it is situated in the general architecture of the mind. This is clearly an open question (unlike the compatibility claim I have just discussed), but one which I am happy to pursue via an elaboration of the model and by confronting the model with empirical and theoretical challenges.

Hudson proposes an alternative model in which

In Adger & Smith (Reference Adger, Smith, Leonie and Karen2005) it is pointed out that the constraints on Buckie was/were variation, specifically the categorical impossibility of was with a third person plural pronominal subject (*they was) is true for present tense be (*they is) and also for main verbs (*they walks). Stating this as memorisation of tokens loses the generalisation that third person plural pronouns do not allow variability in their agreement. This is something which rather needs to be stated as a grammatical fact. I can see how a usage-based approach might do this by extracting a generalisation about agreement and third person pronouns from a set of exemplars, but Hudson provides no argument that this is a BETTER way to approach the data than the analysis I proposed in Adger (Reference Adger2006). It is interesting that Smith, Durham & Fortune (Reference Smith, Durham and Fortune2007) report that children acquiring this variable pattern acquire the categorical impossibility of third person pronouns with variable agreement as soon as they acquire the agreement system in general, and with no analogical leveling. I'd be interested to see how a usage-based account would capture this.Footnote ²

3. The putative flaws

Hudson puts forward a case that the arguments presented in Adger (Reference Adger2006) were flawed in a number of ways. The first part of this case is that middle-aged women have a markedly lower rate of use of we was than the other groups. I suggested in personal communication to Hudson that this was possibly because of prescriptive knowledge. Hudson responds that ‘the trouble with this argument is that it is no stronger than the arguments for excluding prescriptive knowledge from the grammar’ (p. 687). He then says that it is always possible to exclude such knowledge from grammar, but that it is difficult to justify this move linguistically or psychologically. However, the onus is on Hudson here, since he is arguing that a Socio-free account is impossible. I completely agree with him when he writes ‘performance is heavily influenced not only by the subject (you or we) but also by the speaker's age and sex in a complex interactive pattern’ (p. 688) but this does not constitute an argument that age and sex are part of a speaker's grammatical knowledge, as we saw in the previous section.

The second part of the case is that the speakers, once categorised into age and sex groups, rarely show the precise pattern predicted by the mechanisms of Combinatorial Variability specified in Adger (Reference Adger2006). But a major point of Adger (Reference Adger2006) is that this pattern only emerges once enough data is put together. I explicitly discuss the fact that various factors will impact directly on the use of a particular variant in any speech event. My question was whether one could see a general pattern emerging when these factors were controlled for, and my suggestion was that such a pattern could be attributed to the structure of the pool of variants, and hence ultimately to the grammar. I argued that this was precisely what happened when we look at the patterns as a whole (see p. 527, where this is discussed). The lack of enough individual data was precisely the reason that the paper is presented as a plausibility argument, rather than a knock-down argument.

This point is probably worth emphasising here. The grammar G predicts n variants for a particular meaning with a uniform probability distribution. This uniform distribution of variants does not predict a uniform distribution of phonological forms, since the phonological forms are not themselves always evenly distributed over the variants. Moreover, in any particular speech event, the speaker's choice of variant will be given by U, which is sensitive to speaker-internal properties such as intention, processing and memory and to (ultimately internalised) properties of the utterance context, such as who the interlocutor is and what conversation has gone before. Across sociolinguistically categorised groupings we may (or may not) see emergent patterns of higher or lower frequencies for particular forms. Collating the data together is one way to empirically bring out the effect of the uneven distribution of phonological forms over variants predicted by the theory.

The third part of Hudson's case is that I assume that every lexical item has an equal chance of being used; that is, I assume a uniform probability distribution in the set of variants. He claims that common experience shows this is not so (for example, whether you tend to use somebody vs. someone depends on formality). But this argument is backwards. It is an empirical finding that the distribution of someone and somebody is non-uniform, and a departure from the null hypothesis. As is well known in Bayesian probability theory, it is crucial to assume a prior probability distribution and I simply assumed a uniform distribution, the null hypothesis; it would have been a curious move to assume anything else. That this assumption led to the correct predictions is itself an interesting finding.

The final point Hudson makes is a somewhat technical one. He suggests that the feature specification I adopt for singular second person pronouns forces there to be a semantic distinction between generic and personal you. However, this isn't actually the case, since this distinction can be, and I think is, negotiated in the semantic/pragmatic component, with the generic meaning being a pragmatic weakening of the personal one (i.e. it means something like ‘any relevant person’). The idea that φ-features in general are presuppositional and may be manipulated by the pragmatics is, I think, well-established (Cooper Reference Cooper1983, Schlenker Reference Schlenker and Jacqueline2003).