Evaluating Word Representation Features in Biomedical Named Entity Recognition Tasks
Table 5
Performance of CRF-based BNER systems when different types of WR features were used.
System
BioCreAtIvE II GM (%)
JNLPBA (%)
Precision
Recall
-measure
Precision
Recall
-measure
Baseline
87.31
69.20
77.21
71.37
68.68
70.00
Baseline + WR1
86.55
73.18
79.31
70.96
71.44
71.20
Baseline + WR2
87.34
73.91
80.07
71.59
69.55
70.55
Baseline + WR3
86.56
72.22
78.74
71.11
69.88
70.49
Baseline + WR1 + WR2
86.56
75.39
80.59
70.99
71.77
71.38
Baseline + WR1 + WR3
85.77
74.65
79.82
70.77
71.87
71.31
Baseline + WR2 + WR3
87.03
74.90
80.51
71.19
70.41
70.80
Baseline + WR1 + WR2 + WR3
86.54
76.05
80.96
70.78
72.00
71.39
WR1, WR2, and WR3 denote three different types of word representation features: clustering-based, distributional, and word embeddings features, respectively.
