Upper bounds for unsupervised parsing with unambiguous non-terminally separated grammars
Franco M. Luque, Gabriel Infante-López · 2009
Unambiguous Non-Terminally Separated (UNTS) grammars have properties that make them attractive for grammatical inference. However, these properties do not state the maximal performance they can achieve when they are evaluated against a gold treebank that is not produced by an UNTS grammar. In this paper we investigate such an upper bound. We develop a method to find an upper bound for the unlabeled F1 performance that any UNTS grammar can achieve over a given tree-bank. Our strategy is to characterize all possible versions of the gold treebank that UNTS grammars can produce and to find the one that optimizes a metric we define. We show a way to translate this score into an upper bound for the F1. In particular, we show that the F1 parsing score of any UNTS grammar can not be beyond 82.2% when the gold treebank is the WSJ10 corpus.