TABLE 8. Disease Entities With Rare Surface Forms and the Performance of BioBERT With/Without Our Debiasing Method.

Test Entity	COVID-19	47, XXY	African Iron Overload	Bejel	Geographic Tongue	Pinta	PMM2-CDG	Precocious Puberty	VACTERL Association
Frequency	5,287	1,292	39	112	391	124	370	6,903	685
BioBERT	45.7	1.0	6.2	45.5	66.8	50.6	29.0	71.1	31.0
+ Debias.	47.3	2.1	17.4	49.8	76.5	55.0	34.6	90.0	43.8
Improvement	+ 1.6	+ 1.1	+ 11.2	+ 4.3	+ 9.7	+ 4.4	+ 5.6	+ 18.9	+ 12.8