DTranNER: biomedical named entity recognition with deep learning-based label-label transition model

BMC Bioinformatics

Table 2 Performance values in terms of the precision (%), recall (%) and F1-score (%) for the state-of-the-art methods and the proposed model DTranNER

Corpus	BC2GM			BC4CHEMD			BC5CDR-Chemical			BC5CDR-Disease			NCBI-Disease
	P	R	F1	P	R	F1	P	R	F1	P	R	F1	P	R	F1
Att-BiLSTM-CRF (2017)	-	-	-	92.29	90.01	91.14	93.49	91.68	92.57	-	-	-	-	-	-
D3NER (2018)	-	-	-	-	-	-	93.73	92.56	93.14	83.98	85.40	84.68	85.03	83.80	84.41
Collabonet (2018)	80.49	78.99	79.73	90.78	87.01	88.85	94.26	92.38	93.31	85.61	82.61	84.08	85.48	87.27	86.36
Wang et al. (2018)	82.10	79.42	80.74	91.30	87.53	89.37	93.56	92.48	93.03	84.14	85.76	84.95	85.86	86.42	86.14
BioBERT (2019)	85.16	83.65	84.40	92.23	90.61	91.41	93.27	93.61	93.44	85.86	87.27	86.56	89.04	89.69	89.36
DTranNER	84.21	84.84	84.56	91.94	92.04	91.99	94.28	94.04	94.16	86.75	87.70	87.22	88.21	89.04	88.62

Note: The highest performance in each corpus is highlighted in Bold. We quoted the published scores for the other models. For Wang et al. [11], we conducted additional experiments to obtain the performance scores for two corpora (i.e., BC5CDR-Chemical and BC5CDR-Disease) using the software on their open source repository [45]

ISSN: 1471-2105