Garden-Pathing in a Statistical Dependency Parser

This study differentiates between probability models that lead to gardenpathing and those that fail to do so in an incremental dependency parser. Models that take into account intermediate parserstates and part-of-speech pairs correctly reflect human preferences in three wellknown cases: Main Verb vs. Reduced Relative ambiguities, Prepositional Phrase Attachment and Subject-Object ambiguities. Dependency width and direction features were not crucial in these examples, but may ultimately be helpful in accounting for other human sentence processing data. The results support computational proposals about human processing that prioritize stack memory and part-of-speech information over surface distance and dependency direction.