Vine parsing augmented Italian treebanks

This brief article describes our contribution to the EVALITA 2009 Parsing Task, dependency track. The TUT and ISST treebanks are augmented with additional features. MIRA is used to find a weight matrix suited for the Covington algorithm, which is subsequently skewed by discriminatively learned hard constraints on dependency lengths. Our skewed algorithm is linear time and thus asymptotically faster than the cubic time Covington algorithm, but increases in performance are insignificant. Our overall system is thus non-competitive. It is shown, however, that it does contribute significantly to the performance of a more competitive ensemble-based system.