Modeling Morphosyntactic Agreement in Constituency-Based Parsing of Modern Hebrew

We show that naïve modeling of morphosyntactic agreement in a Constituency-Based (CB) statistical parsing model is worse than none, whereas a linguistically adequate way of modeling inflectional morphology in CB parsing leads to improved performance. In particular, we show that an extension of the Relational-Realizational (RR) model that incorporates agreement features is superior to CB models that treat morphosyntax as state-splits (SP), and that the RR model benefits more from inflectional features. We focus on parsing Hebrew and report the best result to date, F₁84.13 for parsing off of gold-tagged text, 5% error reduction from previous results.