Automatic Verb Valency Pattern Recognition.

Penn Treebank corpus. Direct and indirect objects are not explicitly tagged in the corpus since they can be structurally determined, but other potential complements and adjuncts are tagged (e.g., CLR for closely related) (Marcus, Santorini, and Marcinkiewicz, 1993). Using these tags and structural information, the verbs and their valency patterns are extracted for use in a large-scale computational cognitive model of reading (Ball, Heiberg, & Silber, 2007) being developed within the ACT-R cognitive architecture (Anderson, 2007). We have found the patterns of the over 3,000 different verbs used in the Penn Treebank. Further, we have found the most common patterns for different inflections, allowing us to add frequency information for each inflection of each verb to our model. This frequency information is used to set the starting activation of chunks in the reading model in a principled way.