The Differential Sensitivity of Acceptability Judgments to Processing Effects

1 Introduction

Linguists have agreed since at least Chomsky 1965 that acceptability judgments are too coarse-grained to distinguish between the effects of grammatical knowledge (what in Chomsky 1965 would be called competence effects) and the effects of implementing that knowledge (or performance effects). This granularity problem means that for any given putative grammatical phenomenon whose existence is demonstrated by acceptability judgments, it is logically possible that the unacceptability is an epiphenomenon of human language processing. To take a famous example, many works have argued that the island effects delineated in Ross 1967 are due to the processing burdens encountered at island boundaries, and not due to grammatical constraints (e.g., Kluender and Kutas 1993, Kluender 1998, 2004).

With the rise of refined experimental methodologies for collecting acceptability judgments, there has been a renewed interest in identifying the contribution of performance factors—in particular, processing factors—to acceptability judgments. For instance, Fanselow and Frisch (2006) report that local ambiguity in German can lead to increases in the acceptability of ultimately ungrammatical representations if the second possible representation is grammatical. Hofmeister et al. (2007) report that factors affecting the acceptability of Superiority violations also affect the processing of wh-questions as measured in reading times, suggesting that there might be a correlation between processing factors and the acceptability of Superiority violations.

While the picture that emerges from these studies is that acceptability judgments are affected by a wide range of processing effects, this squib presents two experiments that suggest that acceptability judgments are not affected by every processing effect. This differential sensitivity to processing effects suggests a potential evaluation metric for the plausibility of processing explanations: if the proposed processing effect exists independently of the structures under consideration, it should be possible to show that the acceptability effect exists independently of the structure as well.

The experiments reported in this squib build upon one of the major findings of sentence-processing research: the active filling strategy. The active filling strategy is defined by Frazier and Flores d'Arcais (1989:332) as follows: "when a filler has been identified, rank the possibility of assigning it to a gap above all other options." In other words, the human parser prefers to complete long-distance dependencies as quickly as possible. Because the earliest completion site is not always the correct one, the active filling strategy entails the construction of many incorrect temporary representations. Just like their nontemporary counterparts, these temporary representations can be manipulated such that they are either completely grammatical, syntactically ungrammatical, or semantically implausible.

The experiments reported in this squib investigate whether these temporary representations affect the acceptability of the final representation, and if so, which type(s). The results suggest that syntactically ungrammatical temporary representations do lower the acceptability of the final representation, while semantically implausible temporary representations and completely grammatical temporary representations have no effect. This pattern of results suggests (a) that acceptability judgments interact with syntactic violations in a qualitatively different way than semantic violations and pure processing mistakes, and (b) that acceptability judgments are differentially sensitive to effects of sentence processing.

2 Syntactically and Semantically Ungrammatical Representations

In experiment 1, two paradigms that take advantage of the active filling strategy were taken directly from the sentence-processing literature to test the effects of syntactically ungrammatical and semantically ungrammatical temporary representations: the filled-gap paradigm (Crain and Fodor 1985, Stowe 1986) and the plausibility paradigm (Garnsey, Tanenhaus, and Chapman 1989, Tanenhaus, Carlson, and Trueswell 1989). The effects of both paradigms on reading times are so well established in the sentence-processing literature as to serve as standard tools for investigating online construction of filler-gap dependencies.

In the filled-gap paradigm, incremental construction of a wh-dependency proceeds until the processor encounters the first verb, at which point the active filling strategy mandates that the wh-dependency be completed. If the verb has an empty θ-position for the wh-filler, construction of the rest of the representation proceeds as usual as in (1a). If the verb has no empty θ-position, the dependency is still completed, but the following NP receives no θ-role. Thus, a θ-Criterion...