Repetition of relative clause attachments in sentence production: Towards an information-theoretical account of syntactic priming

Christoph Scheepers
Saarland University

chsc@coli.uni-sb.de 

Two sentence completion experiments will be reported in which participants had to generate German equivalents of "the servant of the actress who …" (NP of NP RC) constructions.  Target fragments (where subjects were free to produce either high- or low-attached relative clauses, cf. 1) were preceded by structurally similar prime fragments triggering either high attachment of the relative clause (2a), low attachment of the relative clause (2b), or the generation of a structure that did not contain a relative clause (baseline condition, 2c).

Experiment 1 (showing a baseline bias towards low-attached target RCs) obtained reliably more high-attached target RCs after high attachment primes, but only marginally more low-attached target RCs after low attachment primes.  Experiment 2 employed a different set of items (resulting in a baseline bias towards high-attached target RCs) and obtained significantly more low-attached target RCs after low attachment primes, but only marginally more high-attached target RCs after high attachment primes.  The results are interpreted as evidence for a tendency of language producers to retain sequences of generative rule applications from one trial to the next.  (Note that the results cannot be explained in terms of the context-free rules themselves, as in, e.g., Pickering & Branigan, 1998).  Furthermore, a quantitative model is developed (and backed up by additional data from the literature) that allows for the prediction of priming magnitudes in terms of how much information is conveyed by the relevant primes, given a certain bias in the baseline. The model explains why priming is weaker after primes that are of the same type as the preferred baseline response than after primes of the alternative (non-preferred) type.