Application of attribute grammars to natural language sentence generation

Abstract

The Vinci system generates sentences or phrases in natural languages. The specification of each language has syntactic, lexical and morphological components. The basis of the syntax definition is an attribute grammar which is used to create a tree representing the sentence structure. The tree is subsequently transformed as necessary to obtain related sentence forms, such as the interrogative or passive, and is then subjected to lexical and morphological phases which supply the actual words of the language, and conjugate or decline them. In this paper we are concerned with the attribute mechanism, which is used to ensure agreement between the parts of the sentence. Two types of agreement are sought: grammatical and semantic, and the attributes which control each have slightly different properties. For example, semantic attributes are found to form a partially ordered set. Certain features of natural language are best described by permitting the construction of compound attribute values; others, by the use of non- inherited attributes. Specifications have been written for parts of several languages. The system has been used for experiments in linguistics, and for teaching linguistic theory. Further applications include the generation of exercises for second language learners, and the modelling of errors by second language writers.