Be at Odds? Deep and Hierarchical Neural Networks for Classification and Regression of Conflict in Speech - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 408

For the regression task the Pearson correlation coefficient was selected as
evaluation criterion, being a measure for the prediction quality of continuous-valued
score labels. It is defined as follows:

X

N

1
NO

CC D . y n O /.y n /; (19.25)

nD1

where N is the number of samples, y n represents the prediction for the n -th data
point with target value y n . Further, . O ; O 2 / and .; 2 / are the corresponding means
and variances for the predictions and targets across the data set, respectively.
The Challenge's baseline results were computed adopting a linear kernel SVM
trained using Sequential Minimal Optimization (SMO) (Bishop 2006 ). The SVM
complexity parameter C 2f10 3 ;10 2 ;10 1 ;1g which achieved the best UAR on
the development set was chosen for the reference results and logistic models were
fit to the SVM hyperplane distance based on the training set to obtain (pseudo)
class posteriors. In the Conflict Sub-Challenge, the simple baseline already delivers
a remarkable performance on the binary classification and real-valued regression
tasks as shown in Table 19.4 .

19.6.2 Baseline Features (Feature Set I)

In a first experiment we compared the traditionally used sigmoid hidden units to
ReLu hidden units, the latter being used in combination with dropout training. For
this purpose, we trained standard one-hidden layer neural networks, usually termed
multi-layer perceptrons (MLP), feeding the supra-segmental feature set I as input.
This comprised all 6,373 features as described in Sect. 19.5 . All input features
were normalized to zero mean and unit variance, with the means and standard
deviations being computed on the training set. After normalization the feature set
is approximately Gaussian distributed, which turns out to be beneficial when using
GRBMs.
Using these normalized data as input we trained the MLPs varying the sizes
of the hidden layers and using either traditional sigmoid units or ReLus trained

Table 19.4 Challenge
baseline results

[%] C devel test
CC [score] 0:001 81:6 82:6
UAR [class] 0:1 79:1 80:8
C : complexity parameter in SVM
training (tuned on development set).
devel: result on development set by
training on training set. test: result on
test set by training on the training and
development sets