Table 5 Cross-validation results for BERT-type models.

From: Analysis of the retraining strategies for multi-label text message classification in call/contact center systems

Model name

Reference model

STRATEGY 1

STRATEGY 2

Accuracy

SD

Emotica

SD

Accuracy

SD

Emotica

SD

Accuracy

SD

Emotica

SD

[%]

[%]

[%]

[%]

[%]

[%]

[%]

[%]

[%]

[%]

[%]

[%]

Training data

 BERT

90.92

0.56

71.12

1.81

99.61

0.02

97.96

0.12

98.74

0.14

94.56

0.70

 PolBERT

92.01

0.40

73.74

1.63

99.86

0.02

99.22

0.18

99.62

0.08

97.96

0.43

 HerBERT

92.35

0.66

75.00

1.91

99.62

0.06

97.96

0.43

98.86

0.47

94.81

1.42

Testing data

 BERT

90.76

0.84

70.87

2.14

93.24

0.60

77.09

2.16

92.06

0.60

74.66

1.64

 PolBERT

91.94

1.2

73.79

4.33

93.61

0.80

78.06

2.03

92.51

0.49

75.05

1.57

 HerBERT

92.01

0.69

74.17

2.17

93.85

0.24

79.13

0.66

93.61

0.90

78.55

2.46

  1. Significant values are in bold.
  2. SD Standard deviation.
Back to article page