Transition-based and graph-based methods for dependency analysis of sentences

Abstract

Dependency grammar is an important tool in semantical analysis of  Hungarian text sources. There are two main approaches in the literature for construction of dependency graphs. One approach is the family of transition-based methods, while the other model uses direct graph construction operations starting from an initial dependency graph. In this paper, we compare these two methods on text sources in Hungarian. We present here our  proposed methods for the implementation of the approaches. For the test evaluation, we use the sentence bank of  UD (Universal Dependency) homepage. Our experiences show that the accuracy depends significantly from the applied method whether it supports only projective grammars or also non-projective structures are involved.  Regarding the rule-based graph transformation approach, we could get initially a very good progress, but later it was very time-consuming to cover all special features of the language.