Speech Generation from Semantic Nets

Abstract

Natural language output can be generated from semantic nets by processing templates associated with concepts in the net. A set of verb templates is being derived from a study of the surface syntax of some 3000 English verbs: the active forms of the verbs have been classified according to subject, object(s), and complement(s); these syntactic patterns, augmented with case names, are used as a grammar to control the generation of text. This text intern is passed through a speech synthesis program and output by a VOTRAX speech synthesizer. This analysis should ultimately benefit systems attempting to understand English input by providing surface structure to deep case structure maps using the same templates as employed by the generator.