ContextAD: Context-Aware Acronym Disambiguation with Siamese BERT Network

Table 3

Experimental results.

Models		Precision	Recall	F1
Baseline (maximum word frequency) [22]		89.00	46.36	60.97
UC3M [10]		92.15	77.97	84.37
Acronym expander [26]		93.57	83.77	88.40
Human performance [22]		97.82	94.45	96.10
BERT-base	Binary classification [16]	91.76	81.60	86.38
	Interpretation combination [29]	89.91	71.64	79.74
	Sentence combination (ours)	93.16	82.43	87.47
	Score fusion (ours)	93.37	82.66	87.69
RoBERTa- base	Binary classification [16]	90.08	76.87	82.95
	Interpretation combination [29]	88.06	54.19	67.09
	Sentence combination (ours)	91.80	78.63	84.70
	Score fusion (ours)	91.99	78.00	84.42
Sci-BERT-base	Binary classification [16]	92.63	85.69	89.02
	Interpretation combination [29]	92.66	81.28	86.60
	Sentence combination (ours)	94.92	88.87	91.80
	Score fusion (ours)	94.96	89.15	91.97

The bold values in the figure represents the best result obtained by the corresponding BERT model under the same conditions.