KoRASA: Pipeline Optimization for Open-Source Korean Natural Language Understanding Framework Based on Deep Learning

Table 2

Parameter comparison between DIET-Base and DIET-Opt (KoRASA).

Parameter	DIET-Base	DIET-Opt (KoRASA)
Epoch	300		500
The number of transformer layers	2		4
Transformer size	256		256
Masked language model	True		True
Drop rate	0.25		0.25
Weight sparsity	0.8		0.7
Embedding dimension	20		30
Hidden layer size	(256, 128)		(512, 128)