|Title:||Degraded Acceptability and Markedness in Syntax, and the Stochastic Interpretation of Optimality Theory|
|Comment:||draft, 5th May 2004, to appear in Fanselow, Fery, Schlesewsky & Vogel (eds.), Gradiencedraft, 5th May 2004, to appear in Fanselow, Fery, Schlesewsky & Vogel (eds.), Gradience in Grammar, OUP, 2005, comments very welcome|
|Abstract:||The paper offers a critical discussion of treatments of gradience in generative syntax, including several proposal made within OT. I present, in a summarised form, novel data from experimental and corpus research on case conflicts in German free relative constructions. The results of these studies show a discrepancy between the results of experiments and corpus research which first of all calls for a reflection on the nature of different empirical methods and which kinds of predictions we expect a model of grammar to make for these methods.
And it calls for the differentiation of grammar-intrinsic and grammar-external factors that influence empirical results. Grammaticality judgement tasks are unnatural in the sense that they force participants to be normative. The frequencies of corpus studies not only reflect how often a particular structure wins a competition (i.e. its probability of winning in a stochastic OT model), but also how often the competition itself takes place. OT models only care about the former, not the latter, and not distinguishing these two factors leads to counterintuitive proposals. However, it can be shown that both kinds of infrequency are guided by markedness.
Critically reviewing attempts to apply Stochastic OT to syntax, I argue that there is a simple and straightforward correlate of gradience (in observations) within OT models: markedness. Winners of OT competitions are grammatical, but nevertheless they are assigned different violation profiles by the grammar. This feature of OT can successfully be correlated with the gradience we find in our empirical studies. Nevertheless, to account for the full range of data, it seems unavoidable to assume that the empirical tasks impose some constraint rerankings on the OT grammar which seem characteristic for these tasks.
Instead of worrying about this result, I take the position that we should expect it, because there is not the one 'innocent' empirical method which reflects the 'pure' grammar. Consequently, we need an 'enlightened' perspective on empirical methods. The predictions of a model should be robust across different empirical methods, but the effects of the methods themselves on their results need to be reflected in the analysis, too.