Abstract

In this talk I will present two advances in graph-based dependency parsing. In the supervised learning setting, I will present a simple cube-pruning framework that allows for the scoring of arbitrary higher-order features with little loss in accuracy over exact methods. Critical to its success is a generalization of the max-violation training strategy of Huang et al. (2012) to approximate search over hypergraphs—of which cube-pruned dependency parsing is a special case. Experiments on a number of languages shows consistent improvements over other state-of-the art parsers with speeds comparable to linear transition-based parsers. Next, I will investigate how to adapt graph-based parsers to a specific target language in a cross-lingual learning setting. Specifically, I will show how features derived from the World Atlas of Language coupled with a novel ‘ambiguity-aware’ self-training algorithm can lead to large improvements in accuracy, especially for target languages that are typologically diverse from the set of source languages. The talk will conclude with some general opinions on the future of dependency parsing.

This is joint work with Hao Zhang (Google), Oscar Täckström (Google), Joakim Nivre (Uppsala), Liang Huang (CUNY) and Kai Zhao (CUNY).