Detecting Speech Interruptions for Automatic Conflict Detection - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 383

Table 18.8 List of the features computed by the mean and correlation functionals
Mean and correlation functionals Feature name

Mean (post ( f N, O g _1 (O))) O1

Mean (post ( f N, O g _2 (O))) O2

Mean (post ( f N, O g _5 (O))) O5

Correlation (post ( f N, O g _1 (O)), post ( f N, O g _2 (O))) O12

Correlation (post ( f N, O g _1 (O)), post ( f N, O g _5 (O))) O15

Correlation (post ( f N, O g _2 (O)), post ( f N, O g _5 (O))) O25

Correlation (post ( f N, O g _1 (O)), post ( f N, O g _2 (O)), post( f N, O g _5 (O))) O125

Mean (post ( f N, L, H g _5 (N))) N5

Mean (post ( f N, L, H g _5 (L))) L5

Mean (post ( f N, L, H g _5 (H))) H5

Table 18.9 Information gain
of the 15 best-ranked LLDs
of the audio feature set,
including the Ov-2 feature set
and the IS-2013 feature set

Features Information gain Rank (/6,428)

Cov_O125_O125 0.43862 1

Cov_O12_O125 0.43758 2

Cov_O1_O12 0.43586 3

Cov_O1_O125 0.43177 4

Cov_O15_O125 0.42914 5

Cov_O12_O12 0.42858 6

Cov_O25_O125 0.41965 7

Cov_O12_O15 0.41957 8

Cov_O12_O25 0.41431 9

Cov_O15_O15 0.41429 10

Cov_O15_O25 0.41325 11

Cov_H5_O1 0.40915 12

Cov_H5_O15 0.40849 13

Cov_O1_O25 0.40705 14

Cov_H5_O12 0.40509 15

Two overlap feature sets have been defined. The first feature set, called Ov-1,
consisted of 28 features; it was computed by the covariance functional applied to the
features that are related to the f N, O g detectors (O1, O12, O15, O2, O25, and O125).
The second feature set, called Ov-2, consisted of 55 features; it was computed by
the covariance functional applied to the features that are related to the f N, O g and
f N, L, H g detectors (O1, O12, O15, O2, O25, O125, N5, L5, and H5). These two
feature sets will allow a contrastive test to measure the contribution of the f N, L,
H g _5 detector in the detection of conflict. The method of information gain was used
to analyze the feature relevance of the Ov-2 set in comparison with those of the
IS-2013 set. Table 18.9 gives the information gain computed on the Train set and
the rank on 6,428 features (55 features from the Ov-2 set and 6,373 features from the
IS-2013 set) of the most relevant features for the conflict detection. The best feature
is the Cov_O125_O125 feature (which is equal to O125 multiplied by O125).