Clearing the garden-path: improving sentence processing through cognitive control training

Abstract

AbstractHow do general-purpose cognitive abilities affect language processing and comprehension? Recent research emphasises a role for cognitive control—also called executive function (EF)—when individuals must override early parsing decisions as new evidence conflicts with their developing interpretation. We tested if training on non-syntactic EF tasks improves readers' ability to recover from misanalysis during language processing. Participants completed pre/post-reading assessments containing temporarily ambiguous sentences susceptible to misinterpretation. Performance increases on a training task targeting conflict-resolution processes (n-back with "lures") predicted improvements in garden-path recovery. N-back responders—those demonstrating reliable training gains—significantly increased their comprehension accuracy across assessments. Their posttest eye-movement patterns also revealed significantly improved real-time revision following entry into disambiguating sentence regions where cognitive control is hypothesised to engage. Untrained participants and n-back non-responders showed no performance changes. The results provide insight into how nonlinguistic functions contribute to parsing and interpretation and suggest that certain language skills are amenable to improvement via domain-general EF training.Keywords: parsingsyntactic ambiguity resolutionlanguage processingcognitive controlexecutive-function trainingworking memoryreading-time measuresconflict/interference resolution AcknowledgementsThe University of Maryland Center for Advanced Study of Language supported this research. The authors thank Sharona Atkins, Jeff Chrabaszcz, Carrie Clarady, Ryan Corbett, Alexei Smaliy, David Alexander and Pooja Datta for their assistance with collecting and scoring data. Additionally, we thank Kiel Christianson, Barbara Forsyth, Seth Greenberg, Jesse Snedeker, Michael Ullman and Scott Weems for valuable feedback on the design of the experiment. Thanks to John Trueswell and Gerry Altmann for key discussion and comments on earlier drafts of the manuscript. Finally, a portion of this work was presented at the 16th Annual Conference on Architectures and Mechanisms for Language Processing (York, England), the 51st Annual Meeting of the Psychonomics Society (St. Louis, Missouri) and the 24th Annual CUNY Conference on Human Sentence Processing (Stanford University, Palo Alto, California). We thank the attendees for their important suggestions.Notes1. Although we subscribe to constraint-based lexicalist perspectives of ambiguity resolution (e.g., MacDonald, Pearlmutter, & Seidenberg, Citation1994; Novick, Kim, & Trueswell, Citation2003; Novick et al., Citation2008; Trueswell & Tanenhaus, Citation1994), testing this theory against serial models (in which individuals start with a syntactically-driven interpretation and revise when needed; Frazier & Fodor, Citation1978) was not the focus of the current experimental efforts, as various constraints were not manipulated to differentiate these models. Moreover, under both accounts, the ambiguous sentences in our experiment should initially lead to an incorrect transitive interpretation, which must be reconciled with conflicting input later in the sentence regardless of how it was developed. We therefore omit discussion of parsing-theory contrasts and describe our materials, as well as readers' processing decisions, in simple terms that do not rely on a particular parsing framework (for a review and theoretical discussion of constraint-based theories with respect to cognitive control, see Novick et al., Citation2005).2. All subjects also completed a third assessment, which occurred three months following Assessment 2 without additional training for the experimental group. Assessments 1 and 2 were of primary interest, as performance at Assessment 2 measured the immediate effects of training vs. Assessment 1 (Assessment 2 was completed approximately one week after trainees finished the regimen). Assessment 3 was included to evaluate maintenance of training effects primarily for the non-syntactic measures of cognitive function (i.e., the assessments of fluid and crystallised intelligence). We do not include Assessment 3 data, as they do not bear on our central hypotheses.3. The four commercial Posit tasks primarily targeted low-level perceptual functions, and were included not because of any expected relation to syntactic ambiguity resolution, but because of theoretical overlap with the other pre-/post-cognitive assessments that subjects completed. The link between the Posit tasks and the other assessments addresses entirely separate research questions beyond the scope of—and unrelated to—the work presented in this paper. We mention them because subjects in the training group completed them during the interval between assessments, but hereafter we limit further discussion and analysis of these tasks because they will be reported in full elsewhere.4. When applying this forced two-cluster approach to n-back performance data, the model identified the same individuals as responders (n = 13) and non-responders (n = 7) as previously categorised, replicating the model-based approach sketched in the main text above. This was also true for Block Span, except that the two subjects identified in the lowest-performing group were clustered here with the other non-responders into a single non-responder group, resulting in the following two clusters: responders (n = 9) and non-responders (n = 10).5. First, we tested if the three groups (responders, non-responders, controls) for each of the remaining three training tasks differed in syntactic ambiguity-resolution performance at Assessment 1. Critically, in mixed-effect models that tested for a Group-by-Sentence-Type interaction at Assessment 1 (similar to what is reported above for the n-back task), only Sentence-Type improved the fit of each model (ps < 0.001). That the Group factor was absent from each best-fitting model (ps > 0.26) indicates no differences in garden-path recovery between responders, non-responders, and controls for LNS, Running Span, and Block Span prior to intervention.