Panelist: Amanda Stent

Recent years have seen huge advances in the availability of interactive speech-enabled interfaces to
databases and in command-and-control. In addition, over the past 20 years researchers have
successfully implemented advanced statistical methods for dialog management and error handling,
based on the seminal work on PARADISE which for the first time enabled parametrization (and
consequently optimization) of dialog systems. These accomplishments may lead some to consider
dialog a “solved problem”. However, we still do not have methods for implementing the famous
“bicycle repair” dialog first introduced by Grosz in 1978 (almost 40 years ago). In particular,
we lack key techniques to handle grounded problem-solving tasks - dialog about tasks that are
grounded in the physical world and involving more actions than “find me information”. Despite
decades of research, we also still lack efficient and accurate techniques for handling key dialog
phenomena illustrated in the “bicycle repair” dialog, in particular understanding and generating
anaphora, and entrainment. What kind of data sets, methods and evaluation metrics will allow us
finally to address the challenges posed by language-rich, situated problem solving interactions
of three or more turns - i.e. the kinds of dialog humans engage in with each other every day?