Institute of Formal and Applied Linguistics

The first part of the talk introduces the transition-based approach to data-driven dependency parsing, where inference is performed as greedy best-first search over a non-deterministic transition system, while learning is reduced to the simple classification problem of mapping each parser state to the correct transition out of that state. The second part of the talk explores three different techniques for handling non-projective trees in transition-based dependency parsing: pseudo-projective parsing, augmented arc transitions, and online reordering. A comparative evaluation based on data from Czech, English and German shows that all three methods can accurately recover non- projective dependencies but with differences in precision and recall that can be related to language-specific properties.