Parsing as classification

Abstract

Dependency parsing can be cast as a classification problem over strings of observations. Compared to shallow processing tasks like tagging, parsing is a dynamic classification problem as no statically predefined set of classes exists and any class to be distinguished is composed of pairs from a given label set (syntactic function) and the available attachment points in the sentence, so that even the number of "classes" varies with the length of the input sentence. A number of fundamentally different approaches have been pursued to solve this classification task. They differ in the way they consider the context, in whether they apply machine learning approaches or not, and in the means they use to enforce the tree property of the resulting sentence structure. These differences eventually result in a different behavior on the same data making the paradigm an ideal testbed to apply different information fusion schemes for combined decision making. © Springer-Verlag Berlin Heidelberg 2010.