Tree Transformations in Data-Driven Dependency Parsing

Tesnière (1959) is often regarded as the founder of the modern theoretical tradition of dependency grammar, but it is not one well-defined theory. Instead, it comprises several related theories, having some core notions in common. Such a notion is the directed relations between pairs of lexical elements, but the formalism and the criteria for this can vary. The primary goal of this study is to investigate different criteria for imposing these relations in a sound way with respect to parsing accuracy. The parser used here is a data-driven dependency parser. The investigated issue here is what kinds of dependency structures are easier to learn and construct for a machine learning method. Memory-based (or instance-based) learning is the machine learning method in focus. In other words, how can the parsing accuracy be increased by making changes in the annotation by means of tree transformations, without losing and distorting information. So what criteria can motivate the directed lexical relations in a dependency based theory? Given that parsing is usually just an intermediate step, the basic question asked is: shall the relations be based on syntactic or semantic criteria? A motivation for using semantic criteria is that they might result in an less complicated extraction semantic information. In several situations more than one solution can be said to fulfill a criterion of being a syntactic relation, and in other situations syntactic and semantic relations may coincide. Issues like these are taken into consideration here. In order to do data-driven, robust dependency based parsing for Swedish, a dependency based treebank containing Swedish sentences is needed. Dependency based treebanks have been developed for several languages, e.g. Prague Dependency Treebank (PDT) (Hajič et al. 2001). The Swedish treebank Talbanken (Einarsson 1976), created in the 70s, has been reconstructed so that different dependency based versions can be extracted (Nilsson et al. 2005). Talbanken, refined using associated reconstruction and tree transformation programs, thus constitutes a suitable data resource for this study.