Toward conversational human-computer interaction

Reviewer: Thomas Portele

The authors of this paper are members of the dialogue group at the University of Rochester, which has been working at the forefront of spoken dialogue research for some years. The paper motivates the importance of spoken interaction with machines, and describes several levels of task complexity that require appropriate underlying techniques to handle the increasingly complex associated phenomena. The authors claim that although general human-like conversational competence is unlikely to be achieved by a machine, for practical tasks with cooperative users, the goal is reachable without much task-specific tuning. They identify some challenges, and describe how these issues are solved in their current system, which deals with emergency calls. They omit any mention of speech recognition performance, which is one of the main bottlenecks for practical applications in real-world environments that include car noises, or mobile-phone channel artifacts. The paper gives a good overview of the work at Rochester, with references to more specific papers. It omits a closer description of some aspects, such as discourse obligations. The text is readable and informative, and demonstrates what dialogue systems will have to offer as soon as a reasonable level of robustness can be achieved. It remains to be shown, by (already planned) field tests, whether the ambitious general framework described in the paper will be able to deliver that robustness level. Online Computing Reviews Service