Going the distance: Memory and control processes in active dependency construction

Abstract

Filler–gap dependencies make strong demands on working memory in language comprehension because they cannot always be immediately resolved. In a series of three reading-time studies, we test the idea that these demands can be decomposed into active maintenance processes and retrieval events. Results indicate that the fact that a displaced phrase exists and the identity of its basic syntactic category both immediately impact comprehension at potential gap sites. In contrast, specific lexical details of the displaced phrase show an immediate effect only for short dependencies and a much later effect for longer dependencies. We argue that coarse-grained information about the filler is actively maintained and is used to make phrase structure parsing decisions, whereas finer grained information is more quickly released from active maintenance and consequently has to be retrieved at the gap site.

Keywords:

• Parsing
• Psycholinguistics
• Syntax
• Memory
• Unbounded dependencies

The authors gratefully acknowledge Shravan Vasishth for helpful analytical advice and discussion.

Supplemental material is available via http://people.ucsc.edu/~mwagers

Notes

¹ We use the terms gap and gap position in a theory neutral way: None of our discussion turns upon whether the tail of an unbounded dependency is an empty category or not.

²McKoon and Ratcliff (1994) object to the reactivation interpretation of cross-modal lexical priming in these sentential contexts on other grounds. They argue, instead, that the goodness-of-fit of the lexical decision target is an important determinant of the reaction time (RT) patterns in these experiments. Thus lower RTs are observed in postverbal positions not because of priming from a reactivated filler, but because the lexical decision target is itself a noun that could occur in the postverbal position. The probe recognition study of McElree (2001) is less vulnerable to this objection, since the probe words were adjectives, and the task was to determine whether a synonymous word was encountered in the prior context.

³Tanenhaus, Stowe, and Carlson (1985); Stowe (1986); Swinney, Ford, Frauenfelder, & Bresnan (1988); Frazier and Clifton (1989); Boland et al. (1995); Pickering, Barton, and Shillcock (1994); Traxler and Pickering (1996); Clahsen and Featherston (1999); McElree (2000); Sussmann and Sedivy (2003); Aoshima, Phillips, and Weinberg (2004); Conklin, Koenig, and Mauner (2004); Lee (2004).

⁴In any experiment of modest complexity, a number of implicit choices are made by the analyst, more than the explicit steps reported in a research report. Simmons, Nelson, and Simonsohn (2011)’s recent study illustrates how such flexibility in the experimental pipeline can lead to increased effective false-positive rates. We have attempted to disclose all major steps in the analysis of the data we present. In the supplementary materials, we supply coefficient tables for the models over untrimmed data. On the first author's web site, we supply the raw data accompanied by R scripts used to process it, so that the reader can replicate or modify the analyses reported here.

⁵Although the design of the current study required that the verbs be distinct from the verbs in Experiment 1, there are several similarities. The verbs in both Experiment 1 and Experiment 2 allow two internal arguments (theme/benefactive in Experiment 1; figure/ground in Experiment 2). Verbs in both experiments allow alternative orderings of the arguments. Finally verbs in both experiments only require one of the arguments to be realized syntactically. Thus the relative complexity of the two classes is plausibly equivalent.

⁶If we took, as an expectation of the effect size, the average difference between filler match conditions in short dependencies, excluding the first word in the direct object region (see Figure 3), we would have a small positive difference corresponding to a within-subjects corrected Cohen's d of 0.33. For a one-sided t-test and a repeated measures design, a sample size of 58 would be required to achieve a power of .80 (1 – β; G*Power 3, Faul, Erdfelder, Lang, & Buchner, 2007).