Degrees of Grammaticality as Evidence for Suboptimal Linguistic Structures

Intuitive judgments of the acceptability of linguistic utterances constitute the primary source of evidence in linguistic theory. However, our understanding of what speakers actually do when they make such judgments is far from complete. Recently, two psycholinguistic studies bearing on this question have become available: Schütze (1996) provides a broad review of the relevant psychological findings and proposes a new model of the judgment process, and Bard et al. (1996) introduce a new experimental methodology for gathering accurate acceptability judgments. The present paperbuilds on both of these new developments, and tries to shed light on one particular aspect of linguistic data: the relevance of degrees of grammaticality to linguistic theory. We use magnitude estimation experiments as proposed by Bard et al. (1996) to obtain graded linguistic judgments and demonstrate how such data can be exploited for testing linguistic hypotheses. As a model for graded grammaticality, we propose an extended version of Optimality Theory, relying on the concept of ranked grammatical constraints that is independently motivated in Optimality Theory. In this approach, the ranking of constraints induces a ranking of grammatical structures, and degrees of grammaticality emerge as a property of suboptimal structures. Degrees of Grammaticality Gradedness and Linguistic Theory The assumption that degrees of grammaticality are relevant to linguistic theory dates back at least to Chomsky (1964), and on an informal level, graded data are regularly used to support linguistic hypotheses (cf. Schütze, 1996 for an extensive discussion). A standard case is the claim that subjacency violations result in only mild deviance, while ECP violations cause strong ungrammaticality. Belletti & Rizzi’s (1988) influential study of psych-verbs builds on this assumption, making use of no less than seven levels of acceptability. However, Belletti & Rizzi’s treatment of graded data is very casual and provides “no general theory of which principles should cause worse violations. The theory makes no prediction about the relative badness of, say, -Criterion versus Case Filter violations, let alone about how bad each one is in some absolute sense. The notion of relative and absolute badness of particular violations is ad hoc, and is used in just those cases where it is convenient” (Schütze, 1996: 43). This seems to be a typical case: even though the existence of graded data and their potential relevance for linguistic research seems to be generally acknowledged, hardly any effort has gone into the theoretical investigation of graded grammaticality, and none of the established grammatical frameworks offers a systematic account of graded data. On the empirical side, proper criteria for the gathering of graded data are lacking, and the use and interpretation of intermediate grammaticality ratings varies greatly between researchers. Gradedness in Extraction Phenomena This study uses extraction from complex NPs as a test case for the investigation of graded grammaticality. Complex NPs are standardly assumed to be islands for extraction. Picture NPs, however, constitute well-known counterexamples to this assumption, as they allow for island violations in certain cases. Kluender (1992) provides a comprehensive survey of the relevant extraction data, explicitly acknowledging its graded nature. Kluender (1992) claims that extractability depends on the determiner of the picture NP and observes that acceptability gradually decreases from (1a) to (1e): (1) a. Who did you see pictures of? b. Who did you see a picture of? c. Who did you see the picture of? d. Who did you see his picture of? e. Who did you see John’s picture of? Kluender attributes this hierarchy to the increase in specificity of the picture NP, and Fiengo (1987) notes that definiteness and singularity contribute to specificity. Extractability also depends on the matrix verb. Kluender gives the following contrasts, where the first verb increases the acceptability of extraction, while the second one reduces it: (2) a. What did John have/analyze a picture of? b. What did John see/criticize a picture of? c. What did John find/discuss a picture of? d. What did John draw/lose a picture of? e. What did John develop/destroy a picture of? Another factor seems to be the specificity of the extracted NP. Evidence comes from extraction from relative clause islands, which Kluender claims becomes worse with decreasing specificity, i.e., from (3a) to (3c): (3) a. Which paper do you really need to find someone you can intimidate with? b. How many papers do you really need to find someone you can intimidate with? c. What do you really need to find someone you can intimidate with? Kluender’s analysis relies on linguistic judgments obtained in the traditional intuitive fashion. It is unclear, however, if the fine-grained acceptability differences claimed by Kluender can be reliably elicited using such an informal methodology, which does not employ a well-defined measurement scale and fails to control for potential experimental biases. As no statistical analysis can be carried out on the data, the acceptability differences cannot be tested for significance. In the following, we will evaluate Kluender’s claims against graded judgment data elicited under experimental conditions. Experimental Evidence