After presenting a novel O(n^3) parsing algorithm for dependency grammar, wedevelop three contrasting ways to stochasticize it. We propose (a) a lexicalaffinity model where words struggle to modify each other, (b) a sense taggingmodel where words fluctuate randomly in their selectional preferences, and (c)a generative model where the speaker fleshes out each word's syntactic andconceptual structure without regard to the implications for the hearer. We alsogive preliminary empirical results from evaluating the three models' parsingperformance on annotated Wall Street Journal training text (derived from thePenn Treebank). In these results, the generative (i.e., top-down) modelperforms significantly better than the others, and does about equally well atassigning part-of-speech tags.
展开▼
