Parsing with Latent Variable Grammars

old_uid9740
titleParsing with Latent Variable Grammars
start_date2011/03/04
schedule11h-13h
onlineno
location_infoplateau E, 3ème étage, salle 3E91
summaryTreebank parsing can be seen as the search for an optimally refined grammar consistent with a coarse training treebank. We describe a method in which a minimal grammar is augmented with latent variables and hierarchically refined using the EM algorithm. The resulting grammars are highly accurate, but vary widely in their underlying representations, depending on their EM initialization point. We use this to our advantage, combining multiple automatically learned grammars into an unweighted product model, without any learning or tuning of combination weights. Despite its simplicity, the the resulting product model gives significantly improved performance over the state-of-the-art individual grammars, and surpasses even discriminative parsing systems on a variety of languages and domains.
responsiblesCrabbé