Tuning DeSR for Dependency Parsing of Italian

DeSR is a statistical transition-based dependency parser that learns from a training corpus suitable actions to take in order to build a parse tree while scanning a sentence. DeSR can be configured to use different feature models and classifier types. We tuned the parser for the Evalita 2011 corpora by performing several experiments of feature selection and also by adding some new features. The submitted run used DeSR with two additional techniques: (1) reverse revision parsing, which addresses the problem of long distance dependencies, by extracting hints from the output of a first parser as input to a second parser running in the opposite direction; (2) parser combination, which consists in combining the outputs of different configurations of the parser. The submission achieved best accuracy among pure statistical parsers. An analysis of the errors shows that the accuracy is quite high on half of the test set and lower on the second half, which belongs to a different domain. We propose a variant of the parsing algorithm to address these shortcomings.