David Traum, Michael Fleischman, and Ed Hovy
Natural language generation is a broad field, given the wide variety of different applications for text generation. Perhaps one of the most challenging of these applications is natural language generation for spoken dialogue systems. In spoken dialogue systems, real-time throughput is required, which constrains the processing to less than a second if the system is to seem natural, especially given other processing of input and output. Thus text generation approaches which involve selecting from among many possible alternatives or involve complex calculations to determine preferences is not appropriate. Generation in dialogue is also somewhere in between single-shot sentence generation and generation of extended discourse. On the one hand, single short utterances must be generated because one can not predict a priori exactly how the other dialogue participant(s) will react, and subsequent generation may depend more on the input that is newly provided than any previously available information. On the other hand, dialogues generally have a coherent structure, depending on the goals and overall structure of the task that is being discussed as well as the immediately previous utterance. Thus text-planning notions are still relevant, even if one can not count on being able to produce paragraph-level or longer utterances as pre-planned due to the interactive nature of dialogue.