Improving the recall of biomedical named entity recognition with label re-correction and knowledge distillation

BMC Bioinformatics

Table 3 Comparison of BiLSTM-CRF model results trained on CDWC and CDWA with different re-correction times

Dataset	P (%)	R (%)	F (%)	Dataset	P (%)	R (%)	F (%)
CDR	91.42	83.59	87.86	CDR	91.42	83.59	87.86
CDR + CDWC	90.17	84.49	87.24	CDR + CDWA	94.02	71.02	80.92
CDWC	89.72	83.65	86.58	CDWA	94.75	67.27	78.68
CDWC¹	89.84	89.32	89.58	CDWA¹	90.16	88.94	89.55
CDWC²	90.00	89.35	89.67	CDWA² (CDRA)	91.03	88.31	89.65
CDWC³(CDRC)	89.80	89.82	89.81	CDWA³	90.28	89.03	89.65
CDWC⁴	89.90	89.70	89.80

The highest scores are highlighted in bold
All results are evaluated on the CDR test set. The first two lines are the baselines. For the last 5 lines, each dataset is constructed by the correction model trained with the dataset right above it. The superscript represents the re-correction times. That is, CDWC¹ is the dataset constructed by the correction model trained on the CDWC. The third row datasets are the weakly labeled datasets without re-correction. What’s more, CDWC³ is CDRC, and CDWA² is CDRA

ISSN: 1471-2105