Do listeners use discourse-level information to predict upcoming words in an unfolding sentence? An ERP study

Jos J.A. van Berkum^1,2,3, Colin M. Brown³
Pienie Zwitserlood⁴, Valesca Kooijman³, Peter Hagoort^2,3
berkum@psy.uva.nl, colin.brown@mpi.nl
zwitser@psy.uni-muenster.de, valesca.kooijman@mpi.nl, peter.hagoort@fcdonders.kun.nl
1 Department of Psychology, University of Amsterdam
2 FC Donders Centre for Cognitive Neuroimaging, Nijmegen
3 Max Planck Institute for Psycholinguistics, Nijmegen
4 Westfaelische Wilhelms-Universitaet, Muenster

Everyday conversational experience suggests that a listener can easily and often quite successfully continue somebody else's unfinished sentence. In psycholinguistics, the widespread use of written cloze procedures relies on the fact that if the sentential and wider context is sufficiently constraining, readers also systematically converge on a specific word to continue the sentence with. In spite of these phenomena, many psycholinguists are reluctant to consider the possibility that comprehenders can routinely exploit context to predict or anticipate *specific* upcoming words as a sentence is unfolding. In part, this seems to be due to the frequently made assumption that since language is a generative system, specific and real-time lexical prediction could never be a functional aspect of sentence-level language comprehension. Another likely reason is that in psycholinguistics, the concepts of prediction and anticipation are rather strongly associated with strategies that subjects develop in badly designed experiments.

We believe that neither consideration rules out the routine use of sentential and wider context to predict specific upcoming words as part of normal real-time language comprehension. In our talk, we will present ERP evidence that suggests that listeners do in fact use context in this way. The experiment hinged on (Dutch) stories such as

(1)

The burglar had no trouble locating the secret family vault. Of course, it was situated behind a [...]

which in an off-line written cloze test were predominantly completed with one particular noun (e.g., "painting"). To test whether such partly discourse-based lexical predictions would also occur during real-time spoken-language comprehension, we presented the complete stories in spoken form in an ERP experiment, with one modification: the critical noun was now preceded by a gender-inflected adjective whose inflection either agreed with the anticipated noun's gender

(2)

...behind a big-NEU painting-NEU

or did not agree with it

(3)

...behind a big-COM painting-NEU

Relative to the gender-congruent prenominal adjective, the gender- incongruent adjective elicited a small but reliable ERP effect at the inflection. Because the effect depends on the (arbitrary) lexical gender of an expected but as yet not presented noun, it suggests that discourse-level information can indeed lead people to anticipate specific upcoming words as a spoken sentence unfolds. In addition, it suggests that the syntactic properties of those anticipated words can begin to interact with local syntactic constraints before the word at hand has actually been presented. We will relate these findings to models of language comprehension, and also discuss their relation to 'lexical priming'.