The role of hierarchical structure in syntactic dependency integration

The role of the hierarchical syntactic structure has been controversial in sentence processing: while experience-based models of syntactic expectations or surprisal (Hale, 2001) generally rely on hierarchical sentence structure, memory-based models like the Dependency-Locality Theory (DLT; Gibson, 2000) tend to emphasize the importance of linear distance. Although DLT successfully predicts the difference between object and subject relative clauses in English (Gibson, 2000), it has problems accounting for a similar pattern in Korean (Kwon et al., 2006), where the linear distance between the verb and its object is the same in both constructions. One possible way to remedy this shortcoming of DLT is to replace linear distance by a structural distance measuring the number of syntactic nodes crossed (cf. O’Grady, 1997) when integrating the dependent(s) of a head. We tested to what degree structural distance could replace linear distance by evaluating their relative importance as predictors for reading times in the Dundee Corpus (Kennedy & Pynte, 2005). Both distances were calculated based on dependency relations obtained from parsing the corpus with the Stanford Parser (de Marneffe et al., 2006): for each dependency relation the distance was determined as the number of words (LINEAR DISTANCE) between the head and its dependent and as the number of non-terminal nodes crossed (STRUCTURAL DISTANCE) when traversing the syntactic tree structure from the head to its dependent. First-pass reading times were modeled using linear mixed-effects regressions with subjects and word tokens as random effects. We first fitted a baseline model with the predictors word position, word length, unigram and bigram frequency and all two-way interactions that significantly improved the model in a log-likelihood ratio test. All predictors were centered and scaled to reduce collinearity. We then added linear and structural distance as a predictor to the baseline model: while structural distance yields a reliably better model fit (χ 2 (1, N=189,704) = 14.53, p <.001), linear distance does not significantly improve over the baseline model (