Table 2.

The effect of the input representation on performance.

Input representation	Precision (%)	Recall (%)	F-measure (%)
Worda	47.3	71.7	57.0
Word+positionb	49.1	71.4	58.2
Word+position+POSc	51.6	71.8	60.1
ELMod	57.0	67.4	61.8
ELMo+positione	54.2	74.9	62.9
ELMo+position+POSf	56.3	72.7	63.5
BioBERT+position+POSg	57.9	70.1	63.4

^aThe input representation of the model is the word embedding, which is pretrained by word2vec.

^bThe input representation of the model is the concatenation of the word embedding and position embedding.

^cThe input representation of the model is the concatenation of the word embedding, position embedding, and part of speech (POS) embedding. The F-measure (%) for this representation was an important finding.

^dThe input representation of the model is the deep contextualized word representation.

^eThe input representation of the model is the deep contextualized word representation and position embedding.

^fThe input representation of the model is the deep contextualized word representation, position embedding, and POS embedding. The F-measure (%) for this representation was an important finding.

^gThe word representation is generated from the last hidden layer of the bidirectional encoder representations from transformers for biomedical text mining (BioBERT) [33] in a feature-based approach, which means that the parameters of the BioBERT are not fine-tuned. The input representation of the model is the BioBERT word representation, position embedding, and POS embedding.