ATHAAR dataset aids recognition of medieval Arabic scientific manuscripts

Researchers have developed ATHAAR, a new dataset designed to help vision-language models recognize and transcribe medieval Arabic scientific manuscripts.

The project addresses the difficulty models face with cursive ligatures, varying diacritics, and the scarcity of training data for ancient texts. The team created the first annotation guidelines for transcribing these documents based on textual criticism, a standard method for editing historical texts.

The ATHAAR dataset features manually annotated entries that include different versions of the same text, various calligraphic styles, and content that includes both words and numerals. Testing with modern vision-language models revealed that current systems struggle with the specific characteristics of ancient scientific writing. These results indicate that existing technology requires further refinement to accurately process historical Arabic scripts across different chapters and treatises.