The goal of Natural Language Generation (NLG) is to enable systems to communicate to their users through both spoken and written language. NLG investigates how to automatically produce high-quality natural language text from computer-internal representations of information. The applications of this technology include summarization of stock market data, verbalization of computer-generated mathematical proofs, and dialogue systems. Today, most generation systems rely on massive amounts of linguistic knowledge and manually encoded rules for translating the underlying representation into language. As a consequence, building a robust system requires years of development. As in many other fields of language processing, incorporating statistical methods offers the potential to address this problem.

In my talk, I will first present the traditional architecture of generation systems. Then, I will focus on three novel corpus-based methods for qcontent ordering, lexical choice, and sentence realization.