AI restores Ancient Greek texts with unprecedented accuracy
Summary and headline written by AI from the source article. How we work
Researchers at the Austrian Academy of Sciences developed Apollo Restore, a 24-billion-parameter large language model capable of reconstructing missing text in fragmentary Ancient Greek.
The model excels at “fill-in-the-middle” restoration, meaning it can accurately complete gaps in texts without needing to know how long those gaps are. This represents the first large-scale decoder model for historical Greek, and the first for any ancient Mediterranean language. Evaluations show Apollo Restore surpasses existing models by a considerable margin.
It correctly identifies the appropriate restoration within the top twenty possibilities for 80.6% of documentary papyrus, 54.6% of literary papyrus, and 61.0% of stone inscriptions. A new evaluation method accounting for gap length revealed even greater improvements.
In a blind test, experts in papyrology, the study of ancient manuscripts, epigraphy (the study of inscriptions), and philology (the study of language) overwhelmingly preferred Apollo Restore’s reconstructions to those of competing models, rating it as good as or better than human restorations in 77% of instances. The model even improved readings of PHerc. 1667, a papyrus scroll preserved by the eruption of Mount Vesuvius in 79 CE.
Apollo Restore is part of the Decoding Antiquity initiative, which aims to create specialized large language models for historical languages and manuscripts.
