Predicting individual differences in underspecification: An integrated model of good-enough processing

The theory of good-enough processing (Ferreira et al., 2002) suggests that readers strategically adapt their parsing efforts to task demands and available resources, with the consequence of omitting processes that are less important either for the current goal (Swets et al., 2008) or for building a coherent structure (as postulated by Construal, Carreiras & Clifton, 1993). Indeed, von der Malsburg & Vasishth (2012) have shown that readers with low working memory capacity leave an ambiguous adjunct attachment underspecified more often than high-capacity readers. It is not clear, however, how this adaptation works. Is it necessary to assume that low-capacity readers use a different parsing strategy, or can the difference be explained by a common mechanism? So far there is no detailed model of the good-enough account that could clarify this issue. We propose that underspecification results from a common strategy that aims for an uninterrupted reading process and therefore constrains attachment effort, when possible, by the timing of eye movement control processes. We implemented a model where the process of determining the attachment site of an ambiguous adjunct is subject to this interaction between parsing and eye movements. Specifically, the attachment process gets canceled if it has not reached its final stage by the time the next word has been identified and is ready for integration. A varying behavior due to working memory differences would be predicted by the assumption that low-capacity readers on average take longer to complete the attachment than high-capacity readers, resulting in more cancelations that leave the attachment underspecified. Eye movements were simulated in ACT-R using the eye movement module EMMA (Salvucci, 2001) interacting with a cue-based retrieval parser (Lewis & Vasishth, 2005). Individual differences in working memory capacity were modeled in accordance with previous studies (e.g., Daily et al., 2011) by randomly varying ACT-R goal buffer source activation W (normally distributed with mean 1 and SD 0.25; higher values of W improve speed and accuracy of retrieval processes). Low capacity was defined as W ≤ 1 and high capacity as W > 1. Following Swets et al. (2008), an underspecification is not corrected later in the sentence. If an attachment was made, however, contradicting disambiguation information leads to a repair operation triggering a regression towards the beginning of the sentence. We tested the model predictions against the results of von der Malsburg & Vasishth (2012). In their eye-tracking study of temporarily ambiguous (high vs. low) …