Table 5.

Impact of each component in the token embedding composition in terms of the F1-score (%)

Settings	BC2GM	BC5CDR-Chemical	BC5CDR-Disease	NCBI-Disease
W2V	82.03	92.64	85.17	84.88
ELMo	83.41	93.78	86.76	88.27
ELMo + W2V(=DTranNER)	84.56	94.16	87.22	88.62

Note: “W2V” is a variant model of DTranNER whose embedding layer uses only traditional context-independent token vectors (i.e., Wiki-PubMed-PMC [25]), “ELMo” is another variant model of DTranNER whose embedding layer uses only ELMo, and “ELMo + W2V” is equivalent to DTranNER