Source:
ICGI 2010, Springer, Valencia (2010)
Abstract:
Unambiguous Non-Terminally Separated (UNTS) grammars have good learnability properties but are too restrictive to be used for natural language parsing. We present a generalization of UNTS grammars called Unambiguous Weakly NTS (UWNTS) grammars that preserve the learnability properties. Then, we study the problem of using them to parse natural language and evaluating against a gold treebank. If the target language is not UWNTS, there will be an upper bound in the parsing performance. In this paper we develop methods to find upper bounds for the unlabeled $F_1$ performance that any UWNTS grammar can achieve over a given treebank. We define a new metric, show that its optimization is NP-Hard but solvable with specialized software, and show a translation of the result to a bound for the $F_1$. We do experiments with the WSJ10 corpus, finding an $F_1$ bound of $76.1\%$ for the UWNTS grammars over the POS tags alphabet.
