Svitlana Volkova, Pallavi Choudhury, Chris Quirk, Bill Dolan, and Luke Zettlemoyer
Procedural dialog systems can help users achieve a wide range of goals. However, such systems are challenging to build, currently requiring manual engineering of substantial domain-specific task knowledge and dialog management strategies. In this paper, we demonstrate that it is possible to learn procedural dialog systems given only light supervision, of the type that can be provided by non-experts. The approach works in domains where the required task knowledge exists in textual form, for example as instructional web pages, and the system builders have access to examples of user intent statements, for example from search query logs or dialog interactions. To learn from such textual resources, we describe a new end-to-end approach that first automatically extracts task knowledge from the instructions and then learns a dialog manager that can use this knowledge to provide assistance. Evaluations in a Windows help domain demonstrate that the two components are highly accurate and can be integrated into a system that proved highly effective at helping users achieve their goals.
Publisher Association for Computational Linguistics