. 2023 Mar;60(2):None. doi: 10.1016/j.ipm.2022.103206

Table 5.

Comparison of baseline ( $b a s e$ ) against best performing user setup (+d, +t, +d/t) for each model and dataset. We abbreviate DistilBERT as DBERT in this table. For simplicity, we take the same user setup for both labels ( $f a k e$ / $t r u e$ ). Improvements over base models are underlined.

(a) PolitiFact
fake		P	R	F1
CNN	$b a s e$	.7857	.3333	.4681
CNN	+d	.8235	.4242	.5600

HAN	$b a s e$	.8214	.6970	.7541
HAN	+d/t	.7857	.6667	.7213

DBERT	$b a s e$	.8148	.6667	.7333
DBERT	+d	.8750	.6364	.7368

true		P	R	F1

CNN	$b a s e$	.5000	.8800	.6377
CNN	+d	.5366	.8800	.6667

HAN	$b a s e$	.6667	.8000	.7273
HAN	+d/t	.6333	.7600	.6909

DBERT	$b a s e$	.6452	.8000	.7143
DBERT	+d	.6471	.8800	.7457

(b) GossipCop
fake		P	R	F1
CNN	$b a s e$	.7951	.5542	.6531
CNN	+d	.7882	.5591	.6542

HAN	$b a s e$	.6684	.6502	.6592
HAN	+d/t	.6569	.6650	.6610

DBERT	$b a s e$	.7845	.5468	.6444
DBERT	+d/t	.7724	.5936	.6731

true		P	R	F1

CNN	$b a s e$	.8722	.9552	.9118
CNN	+d	.8732	.9529	.9113

HAN	$b a s e$	.8921	.8988	.8950
HAN	+d/t	.8945	.8910	.8928

DBERT	$b a s e$	.8701	.9529	.9096
DBERT	+d/t	.8811	.9451	.9120

(c) ReCOVery
fake		P	R	F1
CNN	$b a s e$	.7872	.5606	.6549
CNN	+t	.8750	.6364	.7368

HAN	$b a s e$	.7500	.8182	.7826
HAN	+d/t	.7714	.8182	.7941

DBERT	$b a s e$	.7463	.7576	.7519
DBERT	+d/t	.8421	.7273	.7805

true		P	R	F1

CNN	$b a s e$	.8092	.9248	.8632
CNN	+t	.8411	.9549	.8944

HAN	$b a s e$	.9055	.8647	.8846
HAN	+d/t	.9070	.8797	.8931

DBERT	$b a s e$	.8788	.8722	.8755
DBERT	+d/t	.8732	.9323	.9018