Table 4.

Performance evaluation of TrOCR-ctx model on Tabular Data Reconstruction across UoS_Data_Rescue, CORD, SROIE, and PubTabNet datasets. Precision and Recall for table structure recognition are calculated based on an IoU threshold $\ge$ 0.6

	Table structure recognition			Tabular data reconstruction
Dataset	P	R	wF1	Rouge-L	WER	CER	EM	F1 (Char)		F1 (Token)
Without contextual information contextual information (TrOCR)
UoS_Data_Rescue	0.742	0.919	0.805	0.771	0.281	0.254	0.719	0.819		0.719
CORD	0.970	0.715	0.798	0.890	0.043	0.031	0.863	0.890		0.863
SROIE	0.805	0.796	0.785	0.847	0.046	0.039	0.819	0.869		0.819
PubTabNet	0.959	0.814	0.869	0.618	0.584	0.593	0.408	0.525		0.408
With contextual information contextual information (TrOCR-ctx without ByT5 model)
UoS_Data_Rescue	0.742	0.919	0.805	0.778	0.258	0.232	0.742	0.824 ($\mathrm{\Delta }0.61\%$)		0.742 ($\mathrm{\Delta }3.20\%$)
CORD	0.970	0.715	0.798	0.917	0.035	0.023	0.895	0.913 ($\mathrm{\Delta }2.58\%$)		0.895 ($\mathrm{\Delta }3.71\%$)
SROIE	0.805	0.796	0.785	0.872	0.025	0.023	0.875	0.909 ($\mathrm{\Delta }4.60\%$)		0.875 ($\mathrm{\Delta }6.84\%$)
PubTabNet	0.959	0.814	0.869	0.636	0.584	0.593	0.416	0.527 ($\mathrm{\Delta }0.38\%$)		0.416 ($\mathrm{\Delta }1.96\%$)
With contextual information contextual information (TrOCR-ctx with ByT5 model)
UoS_Data_Rescue	0.742	0.919	0.805	0.809	0.245	0.213	0.755	0.850 ($\mathrm{\Delta }3.79\%$)		0.755 ($\mathrm{\Delta }5.01\%$)
CORD	0.970	0.715	0.798	0.917	0.023	0.025	0.914	0.921 ($\mathrm{\Delta }3.48\%$)		0.914 ($\mathrm{\Delta }5.91\%$)
SROIE	0.805	0.796	0.785	0.908	0.023	0.022	0.907	0.914 ($\mathrm{\Delta }5.18\%$)		0.907 ($\mathrm{\Delta }10.74\%$)
PubTabNet	0.959	0.814	0.869	0.640	0.592	0.594	0.426	0.536 ($\mathrm{\Delta }2.10\%$)		0.426 ($\mathrm{\Delta }4.41\%$)