Detecting Speech Interruptions for Automatic Conflict Detection - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 380

Fig. 18.7 Multi-expert architecture scheme of the conflict detector

SVM classifier can be divided into three categories: (1) the kernel methods with
selection of the most appropriate kernel or the creation of a new kernel, (2) the
optimization methods with the addition of constraints, and (3) the sample methods
with data generation or modification of data representations. We have chosen the last
category by developing an SVM-based detector, using as input a composite feature
set. This feature set is a concatenation of selected audio features and posterior-based
features that are computed from the posterior probabilities of the overlap detectors.
The architecture characteristics of this classification system are close to those used
in a mixture of experts (Jordan and Jacobs 1994 ). These approaches have theoretical
advantages, such as a reduction in the hypothesis space and learning consistency. As
described in Fig. 18.7 , the multi-expert architecture scheme of the conflict detector
has consisted of a set of overlap detectors (e.g., X, Y) and a conflict/nonconflict
detector (C). A specialized audio feature set (e.g., X-Feat. set) was associated with
each overlap detector (e.g., X), to represent the utterances. A conflict audio feature
set (Cf-Feat. set) was associated with the conflict detector. This feature set consisted
of the selection of the relevant features (Feat. Select.) that were extracted from the
overlap feature set (Ov-Feat. set) and the IS-2013 feature set (cf. Sect. 18.3.1 ).
A set of functionals (Funct.) was applied to the posterior probabilities of the overlap
detectors (e.g., X-Post and Y-Post) to obtain the Ov-Feat. set.
We chose the overlap detectors giving the best UAR on the Development set
(cf. Table 18.4 and 18.5 ): three two-class (Non-Ov/Ov) SVM-based detectors ( f N,
O g _1, f N, O g _2, f N, O g _5) and one three-class (Non-Ov/LLC-Ov/HLC-Ov) SVM-
based detector ( f N, L, H g _5).

18.5.1 Posterior Probabilities

Logistic regression models (Hosmer and Lemeshow 2000 ) were used to obtain
the posterior probabilities from the four overlap detectors ( f N, O g _1, f N, O g _2,
f N, O g _5, and f N, L, H g _5). These posterior probabilities of the overlap detectors
provide information about the uncertainty of belonging to one class: for example,