Skip to main content
. 2019 Dec 27;20:735. doi: 10.1186/s12859-019-3321-4

Table 3.

Comparison between a series of CLSTM (contextual long short-term memory networks [LSTMs] with conditional random fields [CRF]) experiments and the comparative methods on the NCBI, GM, and CDR corpora using strict matching

Strict matching NCBI GM CDR
Model Trial # p r f p r f p r f
BiLSTM - 78.91 82.60 80.71 72.22 72.44 72.33 83.56 80.26 81.88
BiLSTM-CRF - 82.19 84.58 83.37 80.79 79.81 80.30 87.52 83.58 85.50
GRAM-CNN - 84.45 83.92 84.18 80.23 78.83 79.53 86.08 85.49 85.79
BERT - 81.07 80.73 80.90 81.72 81.59 81.65 86.21 85.23 85.72
CLSTM (word+char levels) 1 84.73 86.67 85.68 81.75 81.14 81.44 87.25 85.66 86.44
2 84.43 85.83 85.12 81.26 80.67 80.96 87.16 85.40 86.27
3 86.18 84.48 85.32 82.07 80.24 81.14 87.93 84.56 86.21
4 85.56 85.21 85.39 82.97 79.66 81.28 87.71 85.17 86.42
5 84.62 85.42 85.02 81.02 80.70 80.86 88.27 84.36 86.27
CLSTM average 85.10 85.52 85.31 81.81 80.48 81.14 87.66 85.03 86.33