GOAL ORIENTED MULTI-SENTENCE TEXT GENERATION

Abstract

To generate text, it ry ,t.em must choose rorne I nformat ion based on 'what to say' from its knowledgebase, decide 'how to organize it', and determine 'how to realize it' In natural language. This work is an attempt to generate coherent text, to achieve communicative goals input by the user, in the domain of Economics. Relations from Rhetorical Structure Theory are used as plans to organize the text as a paragraph tree. The construction of the paragraph tree is a top down goal oriented recursive process. Templates and Definite Clause Grammar are used to generate individual clauses and sentences. They form the components of the Surface Generator which takes over after the paragraph tree has been constructed. The domain knowledgebase Is represented to meet the requirement of generation of natural language. The different components of the Generator : Deep generation, Surface generation and Domain Knowledgebase are independently modifiable. This dissertation discusses the theory, design, implementation of such a generator and evaluates the results. The Generator is written in Turbo :Prolog and runs on a PC XT/AT under MS-DOS.