HYBRID LONG-DISTANCE FUNCTIONAL DEPENDENCY PARSING THESIS presented to the Faculty of Arts of the University of Zurich for the degree of Doctor of Philosophy by Gerold Schneider from Nidau, Switzerland Accepted in the winter semester 2006/2007 on the recommendation of Professor Dr. Michael Hess and Dr. Paola Merlo Zurich, July 2008 i Hybrid Long-Distance Functional Dependency Parsing c 2008 Gerold Schneider
[email protected] [email protected] ii Abstract This thesis proposes a robust, hybrid, deep-syntatic dependency-based pars- ing architecture and presents its implementation and evaluation. The architecture and the implementation are carefully designed to keep search-spaces small with- out compromising much on the linguistic performance or adequacy. The resulting parser is deep-syntactic like a formal grammar-based parser but at the same time mostly context-free and fast enough for large-scale application to unrestricted texts. It combines a number of successful current approaches into a hybrid, comparatively simple, modular and open model. This thesis reports three results: We suggest, implement, and evaluate a parsing architecture that is fast, ro- bust and efficient enough to allow users to do broad-coverage parsing of unrestricted texts from varied domains. We present a probability model and a combination between a rule-based competence grammar and a statistical lexicalized performance disambigua- tion model. We show that inherently complex linguistic problems can be broken down and approximated sufficiently well by less complex methods. In particu- lar (1) on the level of long-distance dependencies, the majority of them can be approximated by using a labelled DG, context-free finite-state based pat- terns, and post-processing, (2) on the level of long-distance dependencies, a slightly extended DG allows us to use mildly context-sensitive operations known from Tree-Adjoining Grammar (TAG), (3) on the base phrase level, parsing can successfully be approximated by the more shallow approaches of chunking and tagging.