Induction of Dependency Structures Based on Weighted Projection

This paper describes a novel weighted projection method of inducing grammatical dependency structures for Polish. Using a parallel English-Polish corpus, the English side is automatically annotated with a syntactic parser and the resulting annotations are projected to Polish via word alignment links. Projected arcs are weighted according to the certainty of word alignment links used in the projection, where arcs projected via intersection links are weighted with the lowest value (corresponding to the highest certainty). Minimum spanning trees induced from such graphs are used to train a parsing model with a publicly available parser-generation system.