Genre-Specific Parsing: Non-Additive Influences of Metrical Stress Pattern on Sentence Processing

Christoph Scheepers and Ralf Rummer

Department of Computational Linguistics and Department of Psychology
Saarland University, Germany

In two self-paced reading experiments, we examined how metrical stress patterns affect the processing and comprehension of temporarily (subject/object) ambiguous sentences in German (see below). The sentences were disambiguated either towards a 'canonical' (subject-first, SO) or a 'non-canonical' (object-first, OS) sequence of grammatical functions. There were three groups of participants in Experiment 1. The first two groups (A and B) were presented with sentences embedded in 'lyrical contexts'; in group A, all canonical (SO) structures implied a metrical
irregularity at the disambiguating noun phrase and all non-canonical (OS) structures were metrically regular; in group B, it was the other way round (metrical irregularity in the OS rather than SO condition). Participants in the third group (C) read the same sentences as in (A), but embedded in 'prosaic contexts' without a predictable metrical pattern. Fillers were
either lyrical and metrically regular (group A and B) or prosaic (group C).
(See http://www.coli.uni-sb.de/~chsc/Lyrics/Metric.pdf for further explication and translations.)

Context (A,B): Die Marktfrau hatte keine Zeit. Der Arzt lief zum Gericht.
Context (C): Die Marktfrau hatte kaum Zeit. Der Arzt ist vor Gericht gegangen.
(A,C) SO: Dass Lisa [?] den Kommissar [acc] erkannte,...
OS: Dass Lars [?] der Kommissar [nom] erkannte,...
(B) SO: Dass Lars [?] den Kommissar [acc] erkannte,...
OS: Dass Lisa [?] der Kommissar [nom] erkannte,...
stand im Schlußbericht.

We hypothesised that metrical irregularity would add some extra processing cost (as measurable in higher RTs) to the 'default' cost known from earlier work on subject-object asymmetries in German. Thus, we expected garden path effects elicited by OS-sequences to be weaker for group (A), but stronger for group (B) - relative to the prosaic control group (C). However, the results draw a strikingly different picture: the OS-garden-path effect was significantly more severe in the prosaic condition (C) than in either of the two lyrical conditions (A or B), which, in turn, did not differ from one another in garden-path strength. Moreover, participants in the lyrical conditions (A,B) performed reliably better on 'who-did-what-to-whom' comprehension questions after each trial than those in the prosaic condition (C). The findings were replicated in Experiment 2, which compared prosaic vs. lyrical sentence processing without inducing any metrical violations. We conclude that reading is not substantially disrupted by metrical irregularity. Rather, a deviation from
the canonical (SO) sequence seems generally more tolerable in lyrical context than in prosaic context, but without resulting in poorer understanding. This suggests that readers employ genre-specific ordering constraints: note that scrambled word orders are quite common in lyrical text (often motivated by keeping a well-formed metrical structure), but very rare in prosaic text. Overall, these results point to the importance of distinguishing between different text genres in order to describe the 'default' human sentence processing behaviour.