Concept recognition as a machine translation problem

BMC Bioinformatics

Table 7 Hardware, memory, and time used for training for all evaluated algorithms

Algorithm	Hardware	Training memory (GBs)	Training time (h)
CRF	CPUs	2–13	1–4
BiLSTM*	GPUs/CPUs**	17	29
BiLSTM-CRF	CPUs	7	15
Char-Embeddings	CPUs	30	84
BiLSTM-ELMo*	GPUs	42	700–1000
BioBERT	GPUs/CPUs**	5	20
UZH@CRAFT-ST BioBERT* [4]	GPUS	120***	200
OpenNMT*	CPUs	620	515
ConceptMapper [20]	CPUs	N/A	N/A

A given training time specifies the total hours if training for all ontology annotation sets were run consecutively, but these can be parallelized by ontology
ConceptMapper runs on CPUs but has no training, as it is a dictionary-based lookup tool, hence the specifications as N/A
*Parallelized per ontology due to time constraints
**Runs significantly faster on GPUs
***Total free RAM available

ISSN: 1471-2105