Table 1 Historians’ performance on epigraphic tasks with varying levels of Aeneas support

From: Contextualizing ancient texts with generative neural networks

Method

Restoration 10-character CER ↓

Province

Date distance ↓

(years)

Confidence ↑

Research start ↑

Parallels added ↑

Top-1 ↑

Top-3 ↑

Onomastics

-

13.7%

23.5%

30.4

-

-

-

Historian

39.0%

27.0%

42.0%

31.3

49.5%

-

-

Historian with Aeneas parallels

33.9%

36.7%

56.7%

21.1

61.1%

75.0%

1.48

Historian with Aeneas parallels and prediction

21.4%

68.3%

78.3%

14.1

70.0%

90.0%

1.58

Aeneas

23.1%

66.7%

73.3%

12.8

-

-

-

  1. Historians’ performance on three epigraphic tasks (restoration, geographical attribution and dating) using 60 inscriptions from the LED test set. Tasks were performed independently, then assisted by Aeneas’ parallels (historian with Aeneas parallels) or by its parallels and predictions (historian with Aeneas parallels and prediction). Metrics include restoration (CER, lower is better), geographical attribution (top-1 and top-3 accuracy), dating (distance in years), historian’s confidence, use of Aeneas’ parallels as research starting points, and the number of parallels used. Arrows (↑ and ↓) indicate the direction of optimal performance for each metric. The highlighted values indicate the top performing method.
Back to article page