Data is Data and Model is Model: You Don't Discard the Data that Doesn't Fit Your Model!

We have read with great interest the article published by Frederick Newmeyer in a recent issue of Language (Newmeyer 2003). As discourse-functional grammarians, we feel compelled to respond to several of N’s arguments, and we also question the nature and use of data in the article. As in his previous publications (e.g. Newmeyer 1998), N misrepresents findings of functionally oriented research and even shows disregard for what we consider standards in the use of data in the field. His discussion is replete with speculation, such as the suggestion that speakers of present-day English would be able to carry on a conversation with Shakespeare¹ (p. 698), and of unsubstantiated claims, such as the claim that in both chess and language, the conscious decision of the player or speaker plays a central role (p. 687). In what follows, however, we limit ourselves to only some of the main issues. We should also say that we find ourselves in agreement with several of N’s points, as we also show below, and that we appreciate N’s willingness to engage in debate with those who disagree with his theoretical position.

There are frequent misrepresentations of functional literature in the article. A brief discussion of §3 may provide a good example of the types of problems that can be found throughout. Regarding the discrepancy between the type of examples generated by classical generative models and actual utterances,² N gives a discussion of Du Bois’s (1987) article on argument structure. N begins by presenting one of Du Bois’s findings, which he words as ‘most utterances consist of a verb with one full argument’ (p. 685, emphasis ours). First of all, this constitutes a misrepresentation, since Du Bois’s claim was not about utterances but about clauses. Second, we are not sure what N means by ‘utterance’. This concept would need to be defined first in order to be used as a category in any kind of scientific inquiry; it is not self-evident what should be considered an ‘utterance’, especially in narrative data such as was used by Du Bois (1987). Even more important, the actual claim by Du Bois was that ‘clauses with zero or one lexical argument are common, [while] clauses with two lexical arguments are rare’ (1987:818); in other words, most clauses in Du Bois’s data consisted of verbs with at most one full NP argument. One might suggest that these are minor wording problems, but since findings in Du Bois 1987 are what N chose to target in his article, which, more significantly, has appeared in the principal journal in the field, it seems important to us to point them out.

N goes on by saying ‘if real speech, the argument goes, is not propositional, then grammars should not be either’ (p. 686). In fact, this is N’s interpretation of the findings of Du Bois 1987 and other work he cites in the section. We do not know who might have made an explicit claim to such an effect, and N refers vaguely to ‘some linguists’ (p. 685); the other studies cited in this section (Fox 1994, Thompson & Hopper 2001) also do not make such a claim. We do think this is a very interesting hypothesis provided by research on actual language use and one to which future research efforts might profitably be directed. It cannot be emphasized enough, however, that one needs to be careful in such an endeavor. That is, even though the statistical findings Du Bois (1987) presents look very interesting, they first need to be replicated a number of times (both within languages and crosslinguistically) before we can feel sufficiently sure that they are robust patterns upon which we can build our theories of human language. We note that research reported in a recent volume by Du Bois, Kumpf, and Ashby (2003) begins to do just what we are suggesting here. Obviously, what counts as sufficient is a question that needs to be established in the field, but we...