Corpus-analysis for NLG

There is a general interest in corpora of human authored texts as a source for acquiring domain knowledge useful for a natural language generation (NLG) system. It is less clear, however, how this can be done in a systematic way. We propose a principled approach towards acquiring domain knowledge through corpus analysis and illustrate its application in the domain of route descriptions. More specifically, we identify different types of knowledge needed in the NLG process and describe a procedure for systematically analyzing a corpus text and for inventorizing these different types of knowledge. We discuss how these procedures fit into a global approach to corpus analysis and into the natural language generation system development cycle.