Linguistically–driven Selection of Correct Arcs for Dependency Parsing

Abstract

Abstract. LISCA is an unsupervised algorithm aimed at assigning a quality score to each arc generated by a dependency parser in order to produce a decreasing ranking of arcs from correct to incorrect ones. LISCA exploits statistics about a set of linguistically–motivated and dependency–based features extracted from a large corpus of automatically parsed sentences and uses them to assign a quality score to each arc of a parsed sentence belonging to the same domain of the automatically parsed corpus. LISCA has been successfully tested on two datasets belonging to two different domains and in all experiments it turned out to outperform different baselines, thus showing to be able to reliably detect correct arcs also representing domain–specific peculiarities.