Is children's reading “good enough”? Links between online processing and comprehension as children read syntactically ambiguous sentences

Abstract

We monitored 8- and 10-year-old children's eye movements as they read sentences containing a temporary syntactic ambiguity to obtain a detailed record of their online processing. Children showed the classic garden-path effect in online processing. Their reading was disrupted following disambiguation, relative to control sentences containing a comma to block the ambiguity, although the disruption occurred somewhat later than would be expected for mature readers. We also asked children questions to probe their comprehension of the syntactic ambiguity offline. They made more errors following ambiguous sentences than following control sentences, demonstrating that the initial incorrect parse of the garden-path sentence influenced offline comprehension. These findings are consistent with “good enough” processing effects seen in adults. While faster reading times and more regressions were generally associated with better comprehension, spending longer reading the question predicted comprehension success specifically in the ambiguous condition. This suggests that reading the question prompted children to reconstruct the sentence and engage in some form of processing, which in turn increased the likelihood of comprehension success. Older children were more sensitive to the syntactic function of commas, and, overall, they were faster and more accurate than younger children.

Keywords:

• Children's sentence processing
• Syntactic ambiguity
• Eye movements
• Reading development
• Good enough sentence processing
• Reanalysis

We thank Rachel Loomes, Paul Forbes, Georgina Bremner, and Nicholas Cooper for research assistance along with the staff and pupils at the primary schools in Oxfordshire who participated in this experiment.

Notes

¹To implement the evaluation of effects at the average of the three levels of the sentence type factor for use with lme4, we first manually calculated two dummy-coded variables to implement the factor and then centred those.

²A potential problem with this measure in this design is that there is evidence that a comma may act like a period, meaning that there will be fewer regressions to the region before the comma for reasons other than the presence of ambiguity (Hirotani et al., 2006). One solution is to use an adjusted measure of go-past duration, which only includes all reading time in the region in question (i.e., all fixations in the region before moving rightwards, but not regressions). We therefore conducted the equivalent analysis with this adjusted measure of go-past: The pattern of results remains the same, with significantly longer go-past durations in garden-path than in comma sentences, only for the postdisambiguating verb region (t = 2.04, p < .05); reading times were longer, but not significantly so, in the disambiguating verb region for garden-path than for comma sentences. Thanks to Steven Frisson for discussion of this point.

³A possible concern is that some of our items had relatively short critical verb regions, which might increase the likelihood of not seeing the effect at that critical verb. However, inspection of the means for go-past showed that the difference between garden-path and comma sentences was numerically larger (rather than smaller) for items with shorter critical verbs: items with critical verb length ≥6, 42 ms difference; items with critical verb <6, 52 ms difference.

⁴Due to an oversight in stimulus creation, in one of our sentences (“While Betty was waking up(,) the neighbours slept soundly”), the postdisambiguating verb region was also the final region of the sentence. To check whether this item was driving this effect, we reran the analysis with it removed. The pattern of results remained identical, with a significant effect of sentence type (t = 3.7, p < .01) only.

⁵Random slopes included in the model were: by-participant slopes for go-past reading time and go-past reading time by sentence-type. In a nonconverging version of the model with maximal random slopes the (negative) main effect of go-past reading time was significant (z = −2.07, p = .04) but the interaction with sentence type was not (z = 1.05, p =.3).

⁶Random slopes included were a by-participant slope for reading time. In a nonconverging version of the model with maximal random slopes the pattern of results was identical to those reported except that the interaction of reading time by verb type was marginal (t = −1.95, p = .05), and the simple effect of reading time for OT verbs was marginal (z = −1.7, p = .08).

⁷Random slopes included in the model predicting reading times were: by-participant slopes for sentence type and sentence type by verb type; by-item slopes for TOWRE score and year group. Random slopes included in the model predicting comprehension were: by-participant slopes for verb type and sentence type and by item slopes for TOWRE, year group, sentence type, and sentence type by year-group. Nonconverging models with full random slopes structures showed an identical pattern of results except that in the model predicting comprehension there was a marginal interaction between TOWRE score and verb type (t = 1.83, p = .07).

⁸Random slopes included in the model were: by-participant slopes for regression and the interaction between regression and sentence type; by-item slopes for regression, the interaction between regression and sentence type, and the interaction between regression and year group. The nonconverging models with full random slopes structures showed an identical pattern of results.

⁹Random slopes included in the model with regressions into the second NP as a predictor were: by-participant slopes for regressions, regressions by sentence type, sentence type and verb type; by-item slopes for regressions and regressions by sentence type. Random slopes included in the model with regressions into the first NP as a predictor were: by-participant slopes for regressions, regressions by sentence type, and sentence type; by-item slopes for regressions, regressions by sentence type, and sentence type. Nonconverging models with full random slope structures showed an identical pattern of results to those reported in the text apart from a reliable interaction between making a regression into the first noun phrase and verb type (z = 2.16, p = .03), which was nonsignificant in the converging model.

¹⁰Random slopes included in the model were: by-participant slopes for question reading time and the interaction between question reading time and sentence type. A nonconverging model with full random slopes structures showed an identical pattern of results to those reported in the text except that the interaction between question reading time and age was not reliable for comma sentences.