Modeling Attachment Decisions with a Probabilistic Parser: The Case of Head Final Structures

We describe an incremental, two-stage probabilistic model of human parsing for German. The model is broad coverage, i.e., it assigns sentence structure to previously unseen text with high accuracy. It also makes incremental predictions of the attachment decisions for PP attachment ambiguities. We test the model against reading time data from the literature and find that it makes correct predictions for verb second sentences; however, the model is not able to account for reading times data for verb final structures because attachment preferences in our training data do not match those determined experimentally. We argue that this points to more general limitations with our type of probabilistic model when it comes to realizing processing strategies that are independent of the data the parsing model is trained on.