Skip to main content

Table 5 Performance comparison of various BERTs

From: Deep learning with language models improves named entity recognition for PharmaCoNER

Method

Mean ± SD

Max

P (%)

R (%)

F1 (%)

P (%)

R (%)

F1 (%)

BERT(Cased)

89.31 ± 0.26

88.00 ± 0.16

88.65 ± 0.12

89.51

88.06

88.78\(^*\)

BERT(Uncased)

89.60 ± 0.81

88.13 ± 0.40

88.86 ± 0.57

90.32

88.65

89.48\(^*\)

NCBI BERT(P+M,Uncased)

89.29 ± 0.67

87.11 ± 0.60

88.18 ± 0.35

89.58

87.30

88.42\(^*\)

NCBI BERT(P,Uncased)

90.20 ± 0.38

88.88 ± 0.52

89.53 ± 0.37

90.76

89.58

90.16\(^*\)

Spanish BERT(Uncased)

89.69 ± 0.74

90.56 ± 0.58

90.12 ± 0.37

90.47

90.72

90.59\(^*\)

Spanish BERT(Cased)

90.42 ± 0.77

90.51 ± 0.69

90.47 ± 0.69

91.76

91.31

91.54

MultiBERT(Cased)

89.53 ± 0.27

89.99 ± 0.43

89.76 ± 0.19

89.75

90.34

90.04\(^*\)

MultiBERT(Uncased)

90.74 ± 0.35

90.39 ± 0.37

90.56 ± 0.25

91.02

90.77

90.89

SciBERT(Bertvoc,Cased)

90.36 ± 0.75

89.55 ± 0.30

89.96 ± 0.40

91.66

89.52

90.58\(^*\)

SciBERT(Bertvoc,Uncased)

91.07 ± 0.71

89.00 ± 0.45

90.02 ± 0.55

91.85

89.36

90.59\(^*\)

SciBERT(Scivoc,Uncased)

90.75 ± 0.86

90.27 ± 0.32

90.51 ± 0.40

92.03

90.28

91.15

SciBERT(Scivoc,Cased)

91.25 ± 0.69

90.30 ± 0.58

90.77 ± 0.40

92.40

89.74

91.05

BioBERTv1.0(+PMC,Cased)

90.54 ± 0.71

89.59 ± 0.31

90.06 ± 0.45

91.09

89.90

90.49\(^*\)

BioBERTv1.0(+P,Cased)

90.44 ± 0.34

89.98 ± 0.64

90.21 ± 0.36

90.75

90.55

90.65\(^*\)

BioBERTv1.0(+P+PMC,Cased)

91.08 ± 0.86

89.76 ± 0.52

90.41 ± 0.42

91.13

90.34

90.73

BioBERTv1.1(+P,Cased)

91.40 ± 0.81

90.90 ± 0.47

91.15 ± 0.60

92.44

91.59

92.01

  1. ‘P’ and ‘M’ denote PubMed and MIMIC-III, respectively. The table is sorted according to the average F1-score, and the highest values are shown in bold
  2. *Significant difference between the means of two models according to the T-TEST statistical test. Specifically, it indicates the model has a significant difference compared with BioBERTv1.1(+P,Cased), with more than 95% confidence interval (\(p<\) 0.05)