Detecting Speech Interruptions for Automatic Conflict Detection - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 371

Fig. 18.4 Percentage of overlap duration as a function of the CSR

Table 18.1 Statistics on the speech duration of the moderator
Moderator—spk-050 Train set Non-Ov Ov LLC-Ov HLC-Ov

Total speech duration (in s) 27,284.7 20,083.5 7,201.2 2,619.0 4,582.2

Speech duration of the moderator (in s) 5,149.7 3,183.7 1,966 1,029.8 936.2

Percentage of the moderator speech duration 18.9 15.9 27.3 39.3 20.4

In the multiparty discussions of the debates, a member of the group held the role
of moderator among the participant members who were taking part in the debate.
We analyzed various statistics that were related to the moderator from the Train set.
In Table 18.1 , the statistics on the speech duration of the moderator are given.
The total speech duration (27,284.7 s) that we accounted for is different from the
total segment duration (23,774 s) of the Train set; it was estimated as the duration of
a segment in which a lonely subject is speaking plus twice the segment duration
in which two subjects are speaking. The speech duration of the moderator was
computed for the various classes of speech: Nonoverlap (Non-Ov) and overlap (Ov)
were split into the two conflict level classes, low-level conflict overlap (LLC-Ov)
and high-level conflict overlap (HLC-Ov). The speech duration of the moderator is
given (in s). The percentage of the moderator speech duration was computed as the
ratio between the speech duration of the moderator and the total speech duration.
From the previous statistics, we note in Table 18.1 that the moderator speaks more
during LLC-Ovs than during HLC-Ovs (39.3 % vs. 20.4 %) and Non-Ovs (39.3 %
vs. 15.9 %).
In Table 18.2 , the modes of the interruptions that were related to the moderator
were analyzed according to the following occurrences: “the moderator interrupted
a participant” or “the moderator was interrupted by a participant.” Clips were
extracted from the video into 30-s duration segments. The mode of interruption that
was related to the moderator was defined from an overlap in which the moderator