Head-driven PCFGs with latent-head statistics

Abstract

Although state-of-the-art parsers for natural language are lexicalized, it was recently shown that an accurate unlexicalized parser for the Penn tree-bank can be simply read off a manually refined treebank. While lexicalized parsers often suffer from sparse data, manual mark-up is costly and largely based on individual linguistic intuition. Thus, across domains, languages, and tree-bank annotations, a fundamental question arises: Is it possible to automatically induce an accurate parser from a tree-bank without resorting to full lexicalization? In this paper, we show how to induce head-driven probabilistic parsers with latent heads from a tree-bank. Our automatically trained parser has a performance of 85.7% (LP/LR F1), which is already better than that of early lexicalized ones.