Over the last thirty years, there has been tremendous progress in the field of Natural Language Generation (NLG). Countless papers have been published on text and sentence planning, lexical realization, lexical choice, and the relation between high-level communicative goals and a variety of NLG subtasks. Yet, summarization systems evaluated in the context of the last two Document Understanding Conferences (DUC) make little or no use of NLG technology. This is particularly disturbing given that many of the DUC participants are NLG researchers who are familiar with the advances in the field. So why do DUC participants choose to use little or no NLG technology? Is it because building good summarizers does not require sophisticated NLG technology? Or is it because NLG does not provide the kind of capabilities summarization systems need?
In this talk, I will try to convince you that high-performance summarizers do require sophisticated NLG technology. And that text summarization can act as a catalyst and help us get rid of the unrealistic assumptions that are at the core of many NLG components.