Abstract

We compare the accuracy of a statistical parse ranking model trained from a fully-annotated portion of the Susanne treebank with one trained from unlabeled partially-bracketed sentences derived from this treebank and from the Penn Treebank. We demonstrate that confidence-based semi-supervised techniques similar to self-training outperform expectation maximization when both are constrained by partial bracketing. Both methods based on partially-bracketed training data outperform the fully supervised technique, and both can, in principle, be applied to any statistical parser whose output is consistent with such partial-bracketing. We also explore tuning the model to a different domain and the effect of in-domain data in the semi-supervised training processes.