Eye-tracking Evidence for Frequency and Integration Cost Effects in Corpus Data

Relative clauses have been investigated extensively in the literature on syntactic processing difficulty. Experimental results (Gibson, 1998; King & Just, 1991) show that English subject relative clauses (SRCs) as in (1a) are more difficult to process than object relative clauses (ORCs) as in (1b). (1) a. The reporter who [R1 attacked] the senator admitted the error. b. The reporter who [R2 the] senator [R3 attacked] admitted the error. These result were obtained for isolated sentences in a fully controlled experimental setting. This raises the question whether the ORC/SRC asymmetry generalizes to naturally occurring, contextualized text. We tested this hypothesis on the Dundee Corpus (Kennedy, 2003), which contains the eyetracking record of 10 subjects reading 50,000 words of newspaper text. Syntactic Prediction Locality Theory (SPLT; Gibson, 1998) explains the ORC/SRC asymmetry based on the cost of integrating the verb's arguments when the verb is reached. ORCs incur a higher integration cost, as more discourse referents intervene between the relative pronoun and the verb than in SRCs. This predicts that the reading time for R3 in (1b) should be higher than that for R1 in (1a). Furthermore, SPLT predicts a higher integration cost on R1 than on R2 because the relative pronoun has to be integrated with the verb in R1 but not with the NP in R2. An alternative explanation is offered by probabilistic accounts that predict shorter reading times for more probable linguistic sequences. For example, McDonald and Shillcock (2003) show that P(w_n|w_n-1), the transitional probability of word w_