Targeted Help for Spoken Dialogue Systems

Targeted Help for Spoken Dialogue Systems: intelligent feedback improves naive users' performance Beth Ann Hockey Research Institute for Advanced Computer Science (RIACS), NASA Ames Research Center, Moffet Field, CA 94035 bahockey@riacs.edu Oliver Lemon School of Informatics, University of Edinburgh, 2 Buccleugh Place Edinburgh EH8 9LW, UK olemon@inf.ed.ac.uk Ellen Campana� Laura Hiatt� Gregory Aist Department of Brain Center for the Study of Language�RIACS and Cognitive Sciences� University of Rochester�and Information (CSLI) NASA Ames Research Center, Stanford University� Rochester, NY 14627� 210 Panama St,�Moffet Field, CA 94035 aist@riacs.edu ecampana@bcs.rochester .edu Stanford, CA 94305 lahiatt@stanford.edu John Dowding RIACS NASA Ames Research Center, Moffet Field, CA 94035 jdowding@riacs.edu James Hieronymus RIACS NASA Ames Research Center, Moffet Field, CA 94035 jimh@riacs.edu Alexander Gruenstein BeVocal, Inc. 685 Clyde Avenue Mountain View, CA 94043 agruenstein@bevocal.com Abstract We present experimental evidence that providing naive users of a spoken dia- logue system with immediate help mes- sages related to their out-of-coverage ut- terances improves their success in using the system. A grammar-based recog- nizer and a Statistical Language Model (SLM) recognizer are run simultane- ously. If the grammar-based recognizer suceeds, the less accurate SLM recog- nizer hypothesis is not used. When the grammar-based recognizer fails and the SLM recognizer produces a recognition hypothesis, this result is used by the Tar- geted Help agent to give the user feed- back on what was recognized, a diag- nosis of what was problematic about the utterance, and a related in-coverage ex- ample. The in-coverage example is in- tended to encourage alignment between user inputs and the language model of the system. We report on controlled ex- periments on a spoken dialogue system for command and control of a simulated robotic helicopter. 1 Introduction Targeted Help makes use of user utterances that are out-of-coverage of the main dialogue system recognizer to provide the user with immediate feedback, tailored to what the user said, for cases in which the system was not able to understand their utterance. These messages can be much more informative than responding to the user with some variant of "Sorry I didn't understand", which is the behaviour of most current mixed initiative di- alogue systems. Providing relevant help messages is a non-trivial problem with mixed initiative sys- tems. There is a much wider range of utterances that the user could sensibly say to a mixed initia- tive system at any give point in a dialogue. In ad- dition since the system must determine rather than dictate the dialogue state there is uncertainty about the context in which help needs to be given. Our Targeted Help approach is aimed at addressing this 147 problem using information that can reasonably be extracted from imperfect input. To implement Targeted Help we use two rec- ognizers: the Primary Recognizer is constructed with grammar-based language model and the Sec- ondary Recognizer used by the Targeted Help module is constructed with a Statistical Language Model (SLM). As part of a spoken dialogue sys- tem, grammar based recognizers tuned to a do- main perform very well, in …

Content truncated to 500 words