Understanding underspecification: A comparison of two computational implementations

Abstract

Swets et al. (2008. Underspecification of syntactic ambiguities: Evidence from self-paced reading. Memory and Cognition, 36(1), 201–216) presented evidence that the so-called ambiguity advantage [Traxler et al. (1998). Adjunct attachment is not a form of lexical ambiguity resolution. Journal of Memory and Language, 39(4), 558–592], which has been explained in terms of the Unrestricted Race Model, can equally well be explained by assuming underspecification in ambiguous conditions driven by task-demands. Specifically, if comprehension questions require that ambiguities be resolved, the parser tends to make an attachment: when questions are about superficial aspects of the target sentence, readers tend to pursue an underspecification strategy. It is reasonable to assume that individual differences in strategy will play a significant role in the application of such strategies, so that studying average behaviour may not be informative. In order to study the predictions of the good-enough processing theory, we implemented two versions of underspecification: the partial specification model (PSM), which is an implementation of the Swets et al. proposal, and a more parsimonious version, the non-specification model (NSM). We evaluate the relative fit of these two kinds of underspecification to Swets et al.’s data; as a baseline, we also fitted three models that assume no underspecification. We find that a model without underspecification provides a somewhat better fit than both underspecification models, while the NSM model provides a better fit than the PSM. We interpret the results as lack of unambiguous evidence in favour of underspecification; however, given that there is considerable existing evidence for good-enough processing in the literature, it is reasonable to assume that some underspecification might occur. Under this assumption, the results can be interpreted as tentative evidence for NSM over PSM. More generally, our work provides a method for choosing between models of real-time processes in sentence comprehension that make qualitative predictions about the relationship between several dependent variables. We believe that sentence processing research will greatly benefit from a wider use of such methods.

Keywords:

• Computational modelling
• Underspecification
• Shallow processing

ORCID

Pavel Logačev http://orcid.org/0000-0002-2837-2258

Shravan Vasishth http://orcid.org/0000-0003-2027-1994

We thank Benjamin Swets for sharing the raw data from Swets, Desmet, Clifton and Ferreira (2008), and for providing many helpful and constructive comments on our work, both in the present paper and in Logačev and Vasishth (2015).

Notes

¹Many thanks to Benjamin Swets for providing us with the raw data of the experiment.

²However, all the patterns reported here held true when we applied a stricter exclusion criterion of 8 seconds.

³A possible explanation for why partial specification requires less time than full unambiguous specification is that ambiguous attachments are not semantically interpreted and that establishing one syntactic link and semantically interpreting it requires more time than establishing two syntactic links. However, this explanation, too, requires stipulations about the relative durations of processes.

⁴A further prediction of the non-specification hypothesis is that on non-specification trials in sentences like (1), no information is kept on whether the RC can attach to the general, the assistant, or the CEO.(1) Mary showed the general the assistant of the CEO who was standing on the balcony.

⁵While it is possible to interpret Swets et al.'s proposal such that underspecification occurs in both ambiguous and unambigous sentences, but more frequently in ambiguous sentences, it is not clear why this would be so. Therefore, we will adopt the simplifying assumption that underspecification affects only ambiguous sentences.

⁶There are several explanations for why guesses could be slower than informed decisions: When readers fail to recall the sentence structure, they start searching their memory for clues as to the correct answer. This extensive search in memory could slow down the response. Alternatively, responding to a question could involve a competition between the two response options in a manner similar to the competition-integration model (Spivey & Tanenhaus, 1998; Spivey-Knowlton, 1996). Lack of evidence either way could prolong competition, and thus cause long response times when information about RC attachment is either not present or cannot be recalled. Yet another possibility is uncertainty about past input (Levy, Bicknell, Slattery, & Rayner, 2009); in this case, an incorrect representation of the previously processed material could lead to a retrieval failure due to a retrieval cue not matching the intended target.

⁷An alternative possibility is that RC attachment requires more time when it is carried out during question-answering than when it is carried out during reading. Although it is to be expected that retrieval of the sentence representation will take more time during the question-answering phase than during reading, retrieval is involved in the answering of questions about unambiguous sentences as well. Thus, longer attachment times during question-answering can only be caused by a slowdown in the RC attachment operation after the sentence representation has been retrieved. However, it is not clear what could cause such a slowdown.