Table 5 Experimental results of different PLMs.

From: Toward a stable and low-resource PLM-based medical diagnostic system via prompt tuning and MoE structure

Model (w/o MoE)

All F.

Selected F.

Easy F.

Biological F.

Features

100%

10%

1%

\(100\%\)

\(10\%\)

\(100\%\)

\(10\%\)

\(100\%\)

\(10\%\)

\(100\%\)

\(10\%\)

All

Selected

Easy

Biological

IPDM

\({\textbf {87.75}}\)

\({\textbf {87.10}}\)

\({\textbf {87.50}}\)

86.60

\({\textbf {80.00}}\)

\({\textbf {76.90}}\)

\({\textbf {67.60}}\)

\({\textbf {65.15}}\)

\({\textbf {85.00}}\)

\({\textbf {85.00}}\)

\({\textbf {85.00}}\)

\({\textbf {78.75}}\)

\({\textbf {80.73}}\)

\({\textbf {80.66}}\)

\({\textbf {79.06}}\)

IPDM-[BERT]

87.50

86.40

87.40

86.20

\({\textbf {80.00}}\)

74.75

66.95

64.60

\({\textbf {85.00}}\)

83.75

83.75

78.25

80.50

80.51

78.50

IPDM-[RoBERTa]

87.60

86.70

\({\textbf {87.50}}\)

86.60

79.95

75.05

67.05

\({\textbf {65.15}}\)

83.75

83.25

82.50

78.25

80.66

80.22

78.40

IPDM-[ELECTRA]

87.45

76.05

\({\textbf {87.50}}\)

86.40

78.45

73.40

66.20

63.75

77.50

77.50

82.50

\({\textbf {78.75}}\)

80.22

80.39

78.39

IPDM-[BioBERT]

87.65

86.15

87.40

85.90

79.40

75.10

66.55

64.95

78.75

\({\textbf {85.00}}\)

82.50

\({\textbf {78.75}}\)

80.50

80.45

78.41

IPDM-[SciBERT]

87.30

86.60

87.15

\({\textbf {86.80}}\)

78.10

75.85

66.50

64.30

83.75

83.75

81.25

77.50

80.64

80.59

78.40

  1. The structure of table is consistent with Table 4. All PlMs use only one expert. IPDM achieves the best on every setting, and it has obvious advantages in low resource scenarios.
  2. The optim is marked with bold.