High span readers' use of Gricean (distributional) information predicts individual differences in RC attachment

Ian Hocking and Don Mitchell
I.M.Hocking@exeter.ac.uk, D.C.Mitchell@exeter.ac.uk
University of Exeter

Individual differences in parsing are problematic for theories of syntax. The aims of this poster are to (i) obtain individual differences and (ii) examine their causes. Mendelsohn and Pearlmutter (1999) reported that attachment bias in RC ambiguities could be predicted by 'reading span' (Daneman & Carpenter, 1980) in an off-line study with NP1-PP-NP2-RC structures such as:

(1)

The crowds annoyed the chauffeur of the actor who wanted to go home.

High span readers favoured N2, presumably because their extra capacity makes them aware of the unambiguous Gricean alternative for an N1 attachment. However, other off-line studies of the same structure indicate that high spans favour N1 (Hocking & Mitchell, 2000; Garcia-Orza & Rodriguez, 1999). The present study revisits the same structure with an improved span test (Waters & Caplan, 1996).

In Phase 1, 48 subjects were given a forced-choice off-line questionnaire comprising 64 items (24 experimental, 40 fillers). The experimental items all consisted of NP1-PP-NP2-RC ambiguities. In each case, the preposition was "of":

(2)

The mechanics were modifying the car of the driver who was in the race.

Subjects received either Questionnaire A or B. Sentence (2), from QA, is an example of a congruent NP2-RP relationship. In QB, the NP2-RC relationship was made incongruent by changing the RP to 'that'. This incongruency is distributional rather than definitively syntactic. Hypothesis: high span readers should detect the incongruency and be 'pushed' towards N1 from the default N2, whereas low spans should not. Subjects were also given the Waters and Caplan reading span test (1996). Scores range from 0 (lowest span) to 6 (highest span).

In Phase 2, some weeks later, 71% of the same subjects were given the alternative questionnaire from Phase 1. Thus, if they received QA at Phase 1, they received QB at Phase 2. This was done to counterbalance overall context effects.

A One-way repeated ANOVA was conducted on each questionnaire with subjects as a factor. A significant effect of subject indicates significant individual differences: QA: F(21,504) = 2.35, p = 0.001; QB: F(20,460) = 3.6, p < 0.001. Span did not predict attachment for the N2-RP manipulation. However, reading span regressed against overall NP1 bias gave: R2adj = 0.173, F(1,32) = 7.907, p = 0.008. That is, the higher the span, the more likely the reader to favour N2. This is compatible with high spans' awareness of the non-ambiguous Gricean interpretation of Mendehslon & Pearlmutter (1999). High spans use some kinds of distributional information but not others. Clearly some sources of information are more stable, important or noticeable than others.

References

Daneman, M. & Carpenter P. A. (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behaviour, 19, 450-466.

Garcia-Orza, J. & Rodriguez, J.M. (1999). Processing resources and sentence comprehension. Poster presented in the IV Simposium de Psicolinguistica, Madrid.

Hocking, I. M. and Mitchell, D. (2000). Individual differences in RC attachment: The role of reading-span. Presented at the AMLaP 2000, Leiden, NL.

Mendelsohn, A. & Pearlmutter, N. (1999). Individual differences in reactive clause attachment preferences. Presented at the CUNY Conference on Human Sentence Processing, New York, New York, March 18-20, 1999.