Recovering From Errors in Conversational Dialogue Systems

Spoken dialogue systems can encounter different types of errors, including non-understanding errors. This is where the system realises the user has spoken, but does not understand their utterance. Strategies for dealing with this kind of error have been proposed and tested in the context of slot-filling systems, for example by Dan Bohus with a system which helps reserve conference rooms [1]. However there has been little work into possible strategies for more conversational settings. This dissertation looks at how we could recover from non-understanding errors experienced by a robot tourguide, and tests the strategies in an experimental study. The main hypothesis of this study is that it is beneficial to use strategies which are designed to do something smarter than just asking the user to repeat themselves. The strategies implemented are motivated by the findings of work done on task-based dialogue systems [1,2,3], which suggest it is useful to move the user on through the dialogue instead of getting caught up with the non-understanding error