27 February 2008, Computational Linguistics Seminar, Joakim Nivre (Växjö University and Uppsala University)

Abstract:
This talk summarizes my research on data-driven dependency parsing over the last five years. To put this work into context, I first discuss what it means to parse a sentence in a text (as opposed to the more well understood notion of parsing a sentence with a formal grammar) and propose criteria for the evaluation of text parsers. I then go on to describe my own dependency-based approach to text parsing, which I characterize as "transition-based" (to distinguish it from the other main tradition in data-driven dependency parsing, which I call "graph-based"). In this approach, inference is performed as a 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. I also discuss methods for handling non-projective dependencies (i.e., discontinuous constructions), in particular the widely used pseudo-projective parsing technique, which allows non-projective dependencies to be recovered using a strictly projective parsing algorithm. I conclude with a quick survey of empirical results, focusing on a contrastive error analysis of transition-based vs. graph-based parsing based on data from the CoNLL 2006 shared task.

For more information and abstracts, see http://www.illc.uva.nl/LaCo/CLS/, or contact R.Tsarfaty at uva.nl