Table 5.

The effectiveness of domain-specific contextual word representation according to the mean F1 scores of 30 different random seeds

Pre-trained word model	F1 score
	Mean	SD	Min	Max
PubMed word2vec	53.42	2.51	46.67	56.70
general-purpose ELMo	54.30	3.61	42.76	56.51
random-PubMed ELMo	53.81	3.65	38.89	57.01
specific-PubMed ELMo	55.91	1.49	51.24	57.48

All of the highest scores are highlighted in bold except for the SD. The first-row results derive from the best results of previous experiments (i.e., the last row in Table 4). Note: “PubMed word2vec” denotes the context-free word model, “general-purpose ELMo” denotes the general-purpose contextual word model, “random-PubMed ELMo” denotes the domain-general contextual word model based on 118 million randomly selected tokens abstracts from PubMed, and “specific-PubMed ELMo” denotes the domain-specific contextual word model based on 118 million bacterial-relevant abstracts from PubMed