Table 5 F1 and NED comparison between OCR&UIE and LVLM-based multi-VIE.

From: Visual information extraction from documents via classification-guided large vision-language models

No.

Document type

OCR&UIE-based F1

LVLM-based F1

OCR&UIE-based NED

LVLM-based NED

01

Acad. Qual. Cert.

70.58%

95.90%

0.6512

0.8770

02

Deg. Cert.

65.86%

92.15%

0.6303

0.8617

03

PILPC

51.11%

85.23%

0.6127

0.9022

04

Bus. Lic.

70.85%

94.44%

0.8021

0.9903

05

Soc. Sec. Cert.

43.65%

88.02%

0.3200

0.7635

06

ID Card

91.67%

99.36%

0.8917

0.9821

07

ISO 9001 QMS Cert.

66.67%

97.59%

0.7456

0.9952

08

ISO 14001 EMS Cert.

62.32%

96.30%

0.6546

0.9709

09

SA 8000 Cert.

70.20%

92.67%

0.7092

0.9617

10

ISO 45001 OHSMS Cert.

61.11%

97.12%

0.6873

0.9680

11

CSCRC

69.94%

97.42%

0.7008

0.9754

12

TNSSCC (Risk Assess.)

77.92%

93.47%

0.6474

0.9649

13

TNSSCC (Emerg. Resp.)

69.18%

91.84%

0.6243

0.9480

14

TNSSCC (Des. & Integr.)

75.57%

90.89%

0.7142

0.9221

15

TNSSCC (Sec. Train.)

72.38%

91.43%

0.7123

0.9553

16

PCI

70.34%

94.62%

0.7034

0.9609

Average

68.08%

93.65%

0.67

0.9348