Processing Parallel Structure: Evidence from Eye-Tracking and a Computational Model
Speaker: Jens Apel
Institution: University of Edinburgh
Abstract:
Prior research has shown that the second constituent of coordinate constructions is read faster when its syntactic structure is parallel compared with non-parallel to that of the first constituent (Frazier et al., 1984). Our research addresses whether this phenomenon is restricted to coordination or whether it can occur in other syntactic environments. Frazier et al. (2000) did not find any benefits for parallel compared with non-parallel constructions in non-coordinated sentences. Their evidence, however, derives from self-paced reading. In the present study, we replicate Frazier's experiments using exclusively the eye-tracking paradigm. The results of our experiments have further informed a computational model that predicts reading times in coordinate and non-coordinate constructions.
In the first experiment, participants read 24 German sentences containing a coordinated noun phrase (NP). The internal structure of the coordinated NPs was either parallel (1c) or nonparallel (1a/1b). In the second experiment, the sentences from Experiment 1 were transformed into 24 non-coordinated sentences (2c) by using the same NP pairs. We measured regression path durations: In Experiment 1, we observed faster reading times on the noun of the second NP when the first NP possessed a parallel compared with non-parallel syntactic structure ((1a) 427.8 ms, (1b) 413.8 ms, (1c) 357.4 ms; ps < .05) However, no such parallelism effect was found for the second NP or any other regions in the second experiment.
Since the parallelism effect does not generalize across syntactic environments, it is difficult to accommodate these findings within a general priming account. Frazier et al. (2000) proposed that parallel constructions are most predictable in coordinate structures, and thus faster processed. Based on both this assumption and our findings we developed a computational model. The model uses an incremental, left-corner chart parser, in which the probability of selecting a grammar rule G is raised, if G was already applied earlier in coordination. In addition to accounting for our own findings, the model also accounts for findings of two further experiments on parallelism (Knoeferle & Crocker, 2006).
(1a) Der Panther und der grasende Hase sind vor den Jägern davongelaufen.
"The panther and the grazing rabbit ran away from the hunters."
(1b) Der kraftvolle Panther und der grasende Hase ...
"The powerful panther ..."
(1c) Der lauernde Panther und der grasende Hase ...
"The lurking panther ..."
(2c) Der lauernde Panther beobachtet den grasenden Hasen aus weiter Entfernung.
"The lurking panther watches the grazing rabbit from far away."