Detecting Speech Interruptions for Automatic Conflict Detection - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 372

Table 18.2 Statistics on interruption mode occurrences of the moderator
Moderator—spk-050 Ov LLC-Ov HLC-Ov

# of interruptions 1,353 604 749

# of interruptions by the moderator and
occurrence percentage

645 (47.7 %) 357 (59.1 %) 288 (38.4 %)

# of interruptions of the moderator and
occurrence percentage

198 (14.6 %) 127 (21.0 %) 71 (9.4 %)

was speaking, by examining the previous segment: if in this segment the moderator
was speaking, then the moderator was interrupted by a participant; otherwise, the
moderator interrupted a participant. Taking off the first segment of each clip, an
interruption occurs at the beginning of each overlap segment; the total number
of interruptions in the Train set is 1,353 split into 604 interruptions in LLC-Ovs
and 749 interruptions in HLC-Ovs. The number of interruptions by the moderator
(respectively, the interruptions of the moderator) was computed for the overlaps and
their two categories (LLC and HLC) as well as its percentage of occurrence. We
note that the moderator interrupted the participants more often than the moderator
was interrupted by the participants (47.7 % vs. 14.6 %). Moreover, the moderator
interrupted the participants more in the LLC-Ovs than in the HLC-Ovs (59.1 % vs.
38.4 %).

18.3 Conflict Challenge

The Conflict Challenge was one of the shared tasks that was organized during the
Interspeech 2013 Computational Paralinguistics Challenge (Schuller et al. 2013 ),
which took place from January 15 to May 24, 2013. The task consisted of an
automatic analysis of the group discussions, to retrieve the conflicts. The goal
of this competition was to bridge the gap between research in automatic conflict
detection and the low compatibility of the results. The task data were split into
the Train, Development, and Test sets. The speaker dependence between these sets
was reduced to a minimum that was needed in the real-life settings. As usual,
the criterion to guide the detection strategy is the maximization of the UAR on
the Development set. This set is also used to tune the parameters of the learning
algorithms. Metadata are available only for the Train and Development sets. The
participants did not have access to the labels of the Test set. However, each
participant could upload the instance predictions up to five times, to receive the
confusion matrix and the results from the Test set. The official measure of the
competition is the UAR. An official system of conflict detection was also provided
with the following characteristics: the WEKA data mining tool kit was used as
a framework for the classification task (Hall et al. 2009 ), and the support vector
machine (SVM) classifier with linear kernel and sequential minimal optimization
(SMO) was used for learning; the official set of features (6,373 features), which