Approaches for Learning Constraint Dependency Grammar from Corpora

This paper evaluates two methods of learning constraint dependency grammars from corpora: one uses the sentences directly and the other uses subgrammar expanded sentences. Learning curves and test set parsing results show that grammars generated directly from sentences have a low degree of parse ambiguity but at a cost of a slow learning rate and less grammar generality. Augmenting these sentences with subgrammars dramatically improves the grammar learning rate and generality with very little increase in parse ambiguity.