Incremental natural-language processing with schema-tree adjoining grammars

Abstract

Schema-Tree Adjoining Grammars are proposed for compressing natural-language grammars. Their rules allow for regular expressions associated with inner nodes of elementary trees referring to individual sons of inner nodes in order to describe their occurrences and repetitions in the set of elementary trees specified. In the following, the adequacy of representing a possibly infinite set of trees by one scheme is discussed for incremental natural-language processing where the input is consumed and processed piecemeal. The basic idea is to avoid decisions for which not enough information is available at the moment by the condensed representation of all currently valid alternatives in terms of semi-instantiated schemata. We present an extended Early processing which allows for mixing partially instantiated trees and uninstantiated schemata.