Depth of processing, underspecification, and text change-blindness

Anthony Sanford, Patrick Sturt and Eugene Dawydiak
University of Glasgow

Andrew Stewart
Unilever Research, Port Sunlight

Many experiments in psycholinguistics suggest that interpretation occurs incrementally, and often to a deep, or complete, level. Work in computational linguistics has explored the use of underspecified interpretations, in which the grain of representation is cruder than is possible within various theoretical frameworks. We describe some experiments on failing to detect anomalies that provide evidence for shallow processing - a mechanism supporting the underspecified representations.
The bulk of the talk will concern our recent empirical attempts to examine the level of specification and processing used in the domains of lexical-semantic representation, and plural anaphora resolution using a new technique. We introduce the Text-change-blindness task, and analogue of visual change-blindness, in which failures to detect specific changes in displays may be used to make inferences about the nature of underlying representations.
With the text-change-blindness task, a passage is presented for reading, followed by the same passage with the possibility of a change to one word. For instance a change can be made which is semantically large or small, and which is either in focus or not, due to
sentence (1):

(1) Everyone was wondering what was going on that night. (focus on event)
(1a)Everyone was wondering which man had been arrested. (focus on man)
(2) It was the man with the hat who had been arrested.
(2) The incident appeared in the newspapers.
"Hat" changes to cap (small) or dog (large).

With Experiment 1 we show that:
1. Large semantic changes are more detectable than small ones.
2. Changes to words in focussed positions are more likely to be detected than changes in
unfocussed positions.
3. There is an interaction between these two, suggesting that depth of semantic processing
is greater for words in focussed positions.
These findings are consistent with a focus-driven theory of processing depth.
With experiment 2, we shall present evidence that the representation of individuals as singular or plural is a function of whether the distinction matters for interpretation.
We shall conclude by discussing factors that might modulate the grain of
representations, and why, and with a brief evaluation of the text-change technique as a tool for investigating processing and representation.