The Role of Island Constraints in Second Language Sentence Processing

Abstract

This study investigates whether adult second language learners’ online processing of wh-dependencies is constrained by island constraints on movement. Proficiency-matched Spanish and Korean learners of English completed a grammaticality judgment task and a stop-making-sense task designed to examine their knowledge of the relative clause island constraint and their sensitivity to this constraint in online wh-dependency formation. The results showed that both learner groups have knowledge of the island constraint in English. However, unlike the Spanish speakers, who immediately applied the constraint to prevent the formation of an ungrammatical wh-dependency, the Korean speakers showed evidence of temporarily entertaining an ungrammatical dependency. These findings suggest that the properties of the native language influence the online processing of L2 sentences.

Notes

¹ The signs in the example sentences (e.g., “[ ]”, “___”, “/”) were not presented at the time of testing. The participants saw only the sentences and the period at the end of them. The signs were used here for a clear explanation of the sentence structures and the way in which the sentences were presented.

² An anonymous reviewer suggested that the ungrammaticality of (13b) may arise due to the violation of the subcategorization constraint on the verb saw, resulting in a filled-gap construction at the overt NP the publisher rather than in a violation of the island constraint per se. We acknowledge that the immediate source of ungrammaticality could be because the verb saw cannot take two objects. For the present purposes, however, the filled-gap effect would obtain only if the wh-phrase which book is not assigned a thematic role from the verb wrote in the relative clause. Thus, if a participant correctly rejects the sentence (13b) as ungrammatical based on the fact that the publisher cannot be assigned a thematic role from the verb saw, as suggested by the reviewer, it could be safely assumed that the participant has not interpreted the wh-phrase as the object of the verb within the relative clause, in accordance with the island constraint. If a participant fails to reject the sentence (13b), it may be because either s/he has not applied the island constraint or s/he does not know the subcategorization property of the verb saw. Given the proficiency level of the population under investigation, we conjecture that the former possibility is more likely than the latter.

³ Some experimental sentences had a longer adverbial phrase at the end of the sentence and ended at region 8 or 9. We do not report the results of these regions because not all sentences had these regions.

⁴ An anonymous reviewer duly pointed out that the time taken to make a stop-making-sense judgment should be treated as a reaction time, as it includes both the time taken to read a region and the time taken to make plausibility judgment. We recognize that this makes it more difficult to interpret the reaction time data as compared to more precise measures of reading difficulty such as eye-tracking and self-paced reading. However, as will be presented later, the results show that the pattern of the relationship between the judgment and reaction time data differs between the two L1 groups, suggesting that the two measures do capture different sets of underlying processes that are relevant for revealing L1-induced differences in processing the target structure. In the remainder of the article, we continue to use the term reading times to highlight the different potentials of the two measures that we chose, although we recognize this terminology may be technically imprecise.

⁵ The mean implausibility detection rates and standard errors by each group and condition are presented in numbers in Appendix C.

⁶ We included all the available reading time data in the statistical analyses mainly for methodological considerations. Including only yes response or only no response in the analyses would result in severe imbalance in the number of trials across conditions, undermining the power and reliability of the statistical analyses. In particular, a separate analysis of the no responses would be inappropriate, given the scarcity of no responses in the conditions other than non-island implausible conditions. However, following an anonymous reviewer’s comment that the reading times for the different responses might not reflect the same kind of decision-making process, we conducted an additional analysis of the reading times only for the yes responses in the most crucial region (Region 4), the results of which are reported in footnote 8.

⁷ The mean reading times and standard errors for each group and condition are presented in numbers in Appendix D.

FIGURE 2 Mean reading times of each group in each region.

⁸ As suggested by a reviewer, we conducted an additional analysis of the reading time data for the most critical region (Region 4) with only the trials that received a yes response (see footnote 7). The results showed that the Spanish speakers showed a main effect of plausibility significant by subject, F1(1,20) = 4.377, p < .05; F2(1,19) = 1.717, p > .1, and a significant interaction between plausibility and island, F1(1,20) = 15.31, p < .01; F2(1,19) = 11.71, p < .01. For the Korean speakers, although the main effect of plausibility approached significance, F1(1,29) = 3.975, p = .056; F2(1,19) = 3.618, p = .072, plausibility did not interact with island (Fs < 1). L2 learners, therefore, essentially showed the same pattern of results as in the analyses with all the available data. The only difference from the overall analysis was in the results for the native speakers, who did not show any statistically reliable effects. This null effect might be due to the loss of statistical power resulting from fewer number of analyzable data points for the native speakers as compared to the L2 speakers (due to a higher number of no responses for the native speakers). The native speakers’ results are also consistent with previous studies showing that plausibility or discourse manipulation exerts weaker effects on reading speed for native speakers than for L2 speakers (e.g., Pan & Felser 2011; Roberts & Felser 2011).

⁹ As a reviewer pointed out, the weaker reanalysis effect in the Spanish group and the lack of reanalysis effect in the Korean group may alternatively suggest that the learners’ attempt at reanalysis was less strongly affected by the plausibility of the initial dependency as compared to the native speaker group, irrespective of whether the learners’ reanalysis was ultimately successful. Since this study did not include comprehension questions probing the participants’ final interpretation, it is not possible to directly examine how successful learners ultimately were in reanalysis. We leave this as an open question for future research.

¹⁰ Note, however, that the Korean speakers arrived in the U.S. at a younger age than the Spanish speakers, t(48) = –2.31, p < .05, and the duration of the two groups’ stay in the U.S. was not significantly different, t(48) = 1.131, p > .1.

¹¹ Age of exposure to English and percentage of English use per day were significantly correlated with the L2 learners’ L1. This makes it difficult to interpret the results of overall ANCOVAs on all L2 learners with L1 as a between-participant factor and the experiential variables as covariates because it would be impossible to identify the effects unique to the L1 versus those unique to the experiential variables. For this reason, ANCOVAs were conducted separately for each group. If the amount and quality of L2 experience had exerted a significant effect on the observed pattern of L2 learners’ reading times, then the effect should also be detected within each group.

¹² There were (marginally) significant interactions between the island factor and some of the covariates (island × age of instruction and island × age of immersion for Spanish speakers, island × cloze for Korean speakers; all ps < .056), the reason for which is not clear at this point. These interactions, however, are irrelevant to the major focus of this study.

¹³ Although long-distant scrambling of a wh-element in Korean may result in a long-distance dependency, scrambled sentences are motivated by nonsyntactic factors such as information structure and heaviness of the moved element. Thus, it could be assumed that the syntactic component of the sentence-processing mechanism of Korean speakers may not be as fine-tuned as that of Spanish and English speakers when processing syntactically motivated overt A’-dependencies.

¹⁴ The authors thank the editor for suggesting the ideas of cross-linguistic differences in A’-dependencies and basic word order as potential grounds for the L1 transfer account proposed in the present study.

Additional information

Funding

Funding for this research was provided in part by a National Science Foundation Dissertation Improvement Grant (Linguistics Program; #1022608).