Paper in preparation. Some of the material will be presented at the Tenth Annual CUNY Conference on Human Sentence Processing, 1997.
1) The apartment that the maid who the service had sent over was cleaning every week was well decorated.
2) *The apartment that the maid who the service had sent over was well decorated.
Using an off-line rating task, they found that when the middle VP was removed (as in 2), the resulting construction was rated no worse than the grammatical version (in 1). Gibson & Thomas interpret this as an indication that people find doubly center-embedded relative clause structures just as acceptable when only two verb phrases are included instead of the grammatically-required three.
However, in connectionist simulations involving (among other recursive constructions) sentences with mutiple center-embeddings (Christiansen, in preparation a) the network, having received two verbs, prefers sentence completion (as in 2) over a third verb (as in 1). This predicts that during on-line processing, people would actually prefer the ungrammatical sentence (2) over the grammatical sentence (1).
An on-line, word-by-word grammaticality judgment experiment was carried out to test this prediction using stimuli from Gibson & Thomas (1996). Following the presentation of each sentence (whether accepted or rejected), subjects rated the sentences on a 7-point scale, permitting both rejection data and ratings to be recorded.
An analysis of the pattern of rejections revealed that significantly fewer ungrammatical sentences like 2 (32.4%) were rejected than grammatical sentences like 1 (63.0%) (X^2 = 20.21, p < .001). The ungrammatical sentences were furthermore rated significantly higher than their grammatical counterpars (F1(1, 35) = 15.55, p < .0001; F2(1,5) = 6.85, p < .05).
The results confirm the network predictions. These derive from
intrinsic architectural constraints on processing (also cf. Christiansen & Chater, in submission). In
contrast, symbolic language processing frameworks (e.g., Gibson &
Thomas, 1996) must impose arbitrary memory limitations to comply with
the above results.
Click to request a copy of this paper.
Please email suggestions/errors to mhc27@cornell.edu |