Processing clause-final verbs poses a number of processes to current parsing theories. For working memory based theories focussing on the number and distance of integrations to be made at the verb (e.g. Dependency Locality Theory, Gibson 2000), processing the verb is predicted to become harder the more complements have been encountered beforehand. This effect should be even stronger in subjects with a low reading span. This memory-based view on on-line processing complexity has been questioned by MacDonald and Christiansen (2002), who proposed a model based on Simple Recurrent Networks to predict processing load. Whereas predictions for some constructions strongly resemble those of capacity-based theories, it is yet unclear how processing of clause-final verbs is affected. We ran a series of SRN simulations and on-line reading studies to tackle this question. SRNs were trained with a subset of German, including simple NPnom-verb-(NPdat)-NPacc main clauses and embedded comp-NPnom-(NPdat)-NPacc-verb clauses. About half of the nominative-NPs were modified by a genitive NP. We calculated the grammatical prediction error (GPE) at each word in the test phase. GPEs, and hence estimated reading times, at the clause-final embedded verb were lower in sentences that included a dative NP than in sentences that did not. Including a genitive NP did not reduce the GPE though. The reduction of GPE induced by an additional complement (dative) grew slightly stronger after later training epochs. In summary, our SRN simulations predict an anticipation effect of complements on the verb which gets stronger for more experienced (high span) language processors, and hence the opposite of what DLT predicts. In a series of German self-paced reading and eye-tracking studies on high and low span readers, the prediction of an anticipation effect could be confirmed in major aspects.
Reading time on the embedded verb following more complements (1) was slightly lower than in sentences with the same number of NPs but less verb-complements (2). Adding a pre-verbal adverbial PP (with delight) did not increase reading times on the verb, contrary to DLTs predictions. In sum, the results are in line with anticipation-based accounts like SRNs and contradict integration cost models like DLT.