Are grammatical constraints immune to retrieval interference?

Patil, U. 1 , Vasishth, S. 1 & Lewis, R. 2

1 University of Potsdam
2 University of Michigan

Phillips, Wagers & Lau (2009) have claimed that reflexives in English are immune to interference from structurally inaccessible antecedents because antecedents are retrieved using only structural cues without considering the person, gender and number features. The support for this claim is derived from studies (eg. Nicol & Swinney 1989, Sturt 2003 and Xiang, Dillon, & Phillips 2009) which found either no effect of interference or a late effect. However, the absence of effect in these studies can be attributed to different factors like lack of statistical power, absence of a critical condition and lack of temporal resolution in the methodology. We present a computational model based on the principles of the cue-based retrieval theory (Lewis & Vasishth, 2005) which predicts an interference effect in reflexive binding. We also report an eye-tracking study that confirms the predictions. The simulations predict the interference effect in terms of: (i) processing time at the reflexive, which includes antecedent retrieval time and (ii) percentage of errors in the retrieval of the grammatical antecedent for the reflexive. Prediction (ii) matches the error rates in Sturt (2003) and our web-based replication of the same study. In the eye-tracking study (n=40) we found an early effect of interference from the inaccessible antecedent in terms of first-pass regression probability-- a gender match between the reflexive and the inaccessible NP induced a significantly higher (p=0.04) number of first-pass regressions from the reflexive in the sentence. The effect of a gender match with the grammatical antecedent was observed only in late eye-tracking measures. In sum, this work (i) challenges the claim that the antecedent of a reflexive is accessed using only structural cues, and shows that the interference induced by the intervening noun occurs very early during dependency resolution, and (ii) presents an implemented computational model that predicts the interference effect.