Spoken Dialog Systems for Automated Survey Interviewing

We explore the plausibility of using automated spoken dialog systems (SDS) for administering survey interviews. Because the goals of a survey dialog system differ from more traditional information-seeking and transactional applications, different measures of task accuracy and success may be warranted. We report a large-scale experimental evaluation of an SDS that administered survey interviews with questions drawn from government and social scientific surveys. We compare two dialog confirmation strategies: (1) a traditional strategy of explicit confirmation on low-confidence recognition; and (2) no confirmation. With explicit confirmation, the small percentage of residual errors had little to no impact on survey data measurement. Even without confirmation, while there are significantly more errors, impact on the substantive conclusions of the survey is still very limited.