Verena Rieser and Johanna D. Moore:
Implications for Generating Clarification Requests in Task-oriented Dialogues

In this talk we present a corpus study aimed at informing the on
generation of clarification requests (CRs) in task-oriented dialogues.
First, we present a multidimensional classification scheme for CRs,
consisting of the two meta-dimensions function and form. Using this
scheme, we annotated CRs in human-human dialogues between a travel
agent and customer, available as part of the CMU Communicator corpus.
Second, we compare the results from this study with two previous
corpus studies on CRs, namely the study by Purver, Ginzburg and Healey
on the British National Corpus, and Rodriguez and Schlangen's study on
the Bielefeld corpus. This constellation of corpora annotated for CRs,
allows us to compare CR forms and functions in different types of
corpora, and to draw implications for generating CRs.  Comparing CRs
in the Bielefeld and Communicator corpora to the BNC, we identified
significantly different behaviour for asking CRs in task-oriented
dialogues vs. everyday conversations. Finally, we found a significant
correlation between CR function and surface form across different
types of task-oriented dialogues. This implies that, once the function
of a CR is determined, the generation module can be informed what
surface form it should generate.  We conclude that these results can
be transferred to spoken dialogue systems as long as we know how to
transfer the factors humans consider for setting the function features
of CRs. In current work we explore how to set CR functions for
multi-modal dialogue systems.