A Multichannel Biomedical Named Entity Recognition Model Based on Multitask Learning and Contextualized Word Representations

Table 10

Performance comparison with existing feature-based methods.

Methods	JNLPBA				Methods	NCBI-disease
	Type					Type
Finkel et al. [1]	S	71.6	68.6	70.1	Xu et al. [48]	S	84.8	76.1	80.2
Settles [3]	S	69.1	72.0	70.5	Leaman et al. [49]	S	82.8	81.9	80.9
Yao et al. [18]	S	64.9	76.1	71.0	Dogan et al. [45]	S	83.8	80.0	81.8
Tsuruoka et al. [2]	S	67.5	75.8	71.4	Leaman et al. [50]	S	85.1	80.8	82.9
Tang et al. [6]	S	70.8	72.0	71.4	Limsopatham et al. [22]	S	86.7	81.9	84.3
Chang et al. [4]	S	—	—	71.9	Wei et al. [51]	S	85.3	83.3	84.3
Zhu et al. [20]	S	—	—	72.6	Dang et al. [24]	S	85.0	83.8	84.4
Li et al. [21]	S	74.8	70.9	72.8	Habibi et al. [52]	S	86.4	82.9	84.6
Tsai et al. [53]	S	72.0	74.0	73.0	Zhao et al. [19]	S	85.1	85.3	85.2
Liao et al. [5]	S	72.8	73.6	73.2	Wang et al. [26]	M	85.9	86.4	86.1
Wang et al. [26]	M	70.9	76.3	73.5	Xu et al. [23]	S	86.6	85.8	86.2
Lyu et al. [25]	S	71.2	76.5	73.8	Yoon et al. [27]	M	85.5	87.3	86.4
Li et al. [33]	S	79.6	69.9	74.4	Zhu et al. [20]	S	86.5	88.1	87.3
Gridach et al. [54]	S	74.1	77.7	75.8	Sachan et al. [28]	T	86.4	88.3	87.3
Ours	M	72.6	79.6	76.0	Ours	M	88.2	89.2	88.7

“S” denotes the single-task model. “M” denotes the multitask model. “T” denotes the model based on transfer learning.