Computational Linguistics Colloquium

Monday, 21 May, 16:15
Conference Room, Building C7 4

Evaluating Models of Processing Complexity against Eye-tracking Corpora

We tested the predictions of Dependency Locality Theory (DLT), a theory of linguistic processing complexity, against reading time data extracted from the Dundee corpus, a large corpus of eyetracking data for naturally occurring text. DLT predicts differences in processing complexity for subject and object relative clauses. We found elevated reading times on two distinct regions of these relative clauses, in line with the complexity effects predicted by DLT. We also found that transitional probability has an effect on reading time in these two regions, independent of the DLT effect.

In a subsequent study, we investigated if DLT makes successful predictions for the whole corpus, not just for relative clauses. A significant positive correlation between DLT scores and reading times was only obtained for nouns; for verbs, we found that reading times depend on the type of dependent being integrated, which is not predicted by DLT. We also tested Hale's Surprisal measure against the Dundee corpus, and found a small but significant relationship with reading times. Furthermore, DLT scores and Surprisal are uncorrelated, which indicates that a comprehensive broad-coverage theory needs to model both aspects of sentence processing.

Joint work with Vera Demberg and Roger Levy.

If you would like to meet with the speaker, please contact Ulrike Pado.