Conflict Cues in Call Center Interactions - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 426

Table 20.6 Precision and
recall scores for SVM-RBF2
( N D 6)

Negative (%) Positive (%)

Recall 97.33 33.90

Precision 82.35 80.00

Table 20.7 Best features

Ten best features (PC measure)
1

Speaker role

2 logMelFreqBand_sma[6]_amean

3 logMelFreqBand_sma[7]_amean

4 lspFreq_sma[1]_quartile1

5 logMelFreqBand_sma[1]_percentile99.0

6 logMelFreqBand_sma[6]_quartile1

7 logMelFreqBand_sma[6]_quartile2

8 logMelFreqBand_sma[7]_quartile1

9 logMelFreqBand_sma[7]_quartile2

10 logMelFreqBand_sma[5]_amean

used, i.e., the N best according to PC, and the y -axis shows the accuracy, i.e., the
proportion of the correctly classified instances from the TE. More specifically, the
baseline method achieves a 76.01 % accuracy, while SVM-RBF1 and SVM-RBF2
achieve a maximum accuracy of 78.45 % (for N D 6) and 82.11 % (for N D 6),
respectively. In general, SVM-RBF2 has a higher accuracy for almost all values of
N . This is most probably due to the fact that the speaker role (customer/operator)
feature has the highest PC value.
The effectiveness of the best classifier SVM-RBF2 ( N D 6) was further analyzed
using precision and recall measures for the negative and positive labels. In particular,
as shown in Table 20.6 , the classifier achieves high precision and recall scores for
the negative label. On the other hand, the recall scores for the positive label are much
lower than the negative one.
Table 20.7 includes a ranked list of the ten best features according to PC measure.
A further qualitative analysis of the erroneously classified speech units shows
that the majority of them are attributed to female (79 %) rather than male (21 %)
speakers in accord with the different gender distribution in the corpus. In addition,
operators' speech units are slightly worse classified (53 %) than customers' (47 %)
ones. This is largely due to the nature of the conversations: for example, operators, in
repeated attempts to reassure customers that their problems are being dealt with and
a solution is underway or in explaining a misunderstanding with regard to a certain
procedure, often speak loudly and in a severe manner. This may produce a clash
between their positive intents, as expressed in the verbal content of the utterance
and the paralinguistic properties of their speech that leads to a classification of
these cases as negative instead of positive. The aforementioned cases account for
the 37 % of the errors. On the other hand, there are cases of customers' negative
instances, expressed, however, in a calm and quiet manner; hence, these instances
are erroneously classified as positive.