Table 4.

Results of the syntactic information ensemble on BC5CDR-chemical dataset

Ensemble strategies	Syntactic info.			BioBERT-Base		BioBERT-Large
	PL	SC	DR	F1	$\sigma$	F1	$\sigma$
Baseline	$\times$	$\times$	$\times$	93.50	0.10	93.90	0.31
Sum	$\surd$	$\surd$	$\times$	93.66	0.17	94.20	0.15
	$\surd$	$\times$	$\surd$	93.76	0.16	94.10	0.15
	$\times$	$\surd$	$\surd$	93.81	0.15	94.12	0.14
	$\surd$	$\surd$	$\surd$	93.78	0.25	94.26	0.16
Concatenation	$\surd$	$\surd$	$\times$	93.75	0.23	94.25	0.12
	$\surd$	$\times$	$\surd$	93.80	0.26	94.22	0.16
	$\times$	$\surd$	$\surd$	93.83	0.20	94.31	0.08
	$\surd$	$\surd$	$\surd$	$\mathit{93.88}$	0.26	$\mathit{94.36}$	0.25

The three types of syntactic information used for the ensemble are POS labels (PL), syntactic constituents (SC), and dependency relations (DR). The results are reported in terms of the average F1 scores and the standard deviation ($\sigma$). Sum and concatenation are two ensemble strategies applied to our method