Arabic Fake News Detection: Comparative Study of Neural Networks and Transformer-Based Approaches

Table 2

FND model performance on ArCOV19-Rumors, ANS, and AraNews datasets.

Model \ measure	F1	Precision	Recall	Accuracy
Linear and deep learning (ArCOV19-Rumors)
Linear (WL-W2V)	0.827886	0.831215	0.836915	0.829752
CNN (WL-F)	0.782586	0.881676	0.70797	0.791736
RNN (ChL-F)	0.712743	0.555372	1	0.555372
GRU (WL-W2V)	0.838259	0.833348	0.846716	0.831405
Transformers with gradual unfreezing, special learning rate, and learning rate scheduling (ArCOV19-Rumors)
AraBERT v1-original	0.811362	0.819645	0.83294	0.825083
AraBERT v02	0.669801	0.594531	0.835298	0.628713
AraBERT v2	0.812187	0.846519	0.816502	0.825083
ArAElectra	0.936161	0.929813	0.95675	0.952145
AraGPT2	0.908726	0.914789	0.927801	0.912541
QARiB	0.953345	0.956216	0.956404	0.975248
ArBert	0.891291	0.907135	0.90165	0.914191
MarBert	0.933561	0.950039	0.934968	0.940594
Transformers with a constant learning rate of 1e-5 (ArCOV19-Rumors)
AraBERT v02	0.705171	0.722772	0.69802	0.945545
AraGPT2	0.776255	0.774085	0.806601	0.813531
ArBert	0.953423	0.947729	0.966777	0.958746
Transformers with special learning rate and learning rate scheduling (ArCOV19-Rumors)
AraBERT v02	0.928597	0.960616	0.911881	0.948845
QARiB	0.930478	0.941914	0.92557	0.952145
AraGPT2	0.90882	0.919802	0.927927	0.925743
Transformers with special learning rate and learning rate scheduling (ANS dataset)
AraBERT v02	0.02752	0.06181	0.018186	0.675497
QARiB	0.058888	0.10596	0.045412	0.688742
AraGPT2	0.203931	0.278146	0.191023	0.642384
Transformers with special learning rate and learning rate scheduling (AraNews dataset)
AraBERT v02	0.886098	0.800455	0.999121	0.800264
QARiB	0.887142	0.800422	0.999093	0.800264
AraGPT2	0.659546	0.509349	1	0.509349