Information Technology Reference
In-Depth Information
in courts (Atkinson and Drew
1979
), in political interviews (Beattie
1982
), in
press conferences (Schegloff
1987
), in mediation (Garcia
1991
), in professional
meetings (Boden
1994
), in talk shows in which interpersonal conflicts are expressed
(Brinson and Winn
1997
), in auctions (Heath and Luff
2007
), in political debates
(Valente and Vinciarelli
2010
), and in political meetings that involve large groups
of people in which everyone can contribute ideas, opinions, and proposals and in
which opposition is also expressed (Mondada
2013
). The role of the chair has
been analyzed in various studies (Boden
1994
; Svennevig
2008
; Mondada
2012
).
Prediction of the speaker order in turn-taking was investigated in news, talk shows,
and meetings (Barzilay et al.
2000
; Vinciarelli
2009
).
18.1.2
Guidelines and Overview
In related work on conflict detection in conversational interactions (Valente and
Vinciarelli
2010
; Pesarin et al.
2012
), turn-taking patterns and overlaps between
speakers are shown to be informative with respect to classification into the presence
or absence of conflict. The total amount of overlap and the minimum pitch during
overlap were found to be the features that correlated the most with conflict (Kim
et al.
2012c
). A widely adopted classification of interruptions/overlaps is collabora-
tive or competitive in reference to the “cooperative-competitive” dimension of the
conflict-handling style. While communication strategies are naturally collaborative,
this preponderance is not the case for conflict dialogues, in which competitive
strategies are the norm. The detection of competitive interruption is a difficult
problem in relation to the search of the TRPs. Spectral content and intonation
contour are not sufficient to locate these places. Furthermore, the perception of
the conflict can be different in the case of the constrained organization of turn-
taking, such as institutional interactions (interview, debate, meeting). Competitive
strategies such as those of the moderator or the chairman appear to be natural in
this context and are not perceived as conflicting. Our experiments relate to the
classification of audio clips into two classes of conflict level (low and high) during
the Interspeech 2013 Conflict Challenge. The clips, which were extracted from
political debates, have been annotated into conflict levels, using crowdsourcing
to model the perception of the people. For our design of the conflict detector,
we categorized the overlapping speech into low- and high-level conflict overlap.
We made the assumption that these categories can be detected from acoustic cues.
We focus our study on a multi-resolution framework for the detection of the overlaps
and a multi-expert architecture to include knowledge about overlap in the automatic
conflict detector.
This chapter is organized as follows: Section
18.2
presents the speech material
that we used for the experiments on conflict detection; it describes and analyzes
the statistical characteristics of the corpus while focusing on interruptions and the
moderator's role. Section
18.3
describes the Conflict Challenge and the vari-
ous audio feature sets that were used for our investigations. In Sect.
18.4
,the
Search WWH ::
Custom Search