Parsing with Underspecifications

This paper describes a direct parser for Schema–Tree Adjoining Grammars (S–TAG) which explores schemata, i.e. underspecified elementary rules. Basically, a schema in a S–TAG represents a possibly infinite set of elementary rules by folding up all actual substructures and depicting them in terms of a regular expression (RX). Hence, S–TAGs provide a more condensed grammar representation. In the following we depict a direct parser which exploits regular expressions, i.e. underspecified rules. The parser is based on the TAG parser by Yves Schabes — S–TAGs are closely related to Tree Adjoining Grammars as outlined in this paper. Our parser itself reuses Earley’s entire idea in order to parse the regular expression denoting tree sets. Consequently, the schemata remain partially underspecified as long as no evidence from the input determines a concrete substructure. Basically, less items are predicted and completed if schemata are inspected. However, the worst–case behaviour remains the same as for ordinary TAGs. Finally, the integration of the parser in the project INTEGENINE is outlined. We apply the parser as a bidirectional generation component in a text generation system.