Can machines read the cuneiform tablets that no scholar has read?
Five answers, none aheadNone is ahead yet. See what would settle it.
Where it stands
Five answers have been proposed. None is ahead on the evidence yet. Five are still possible.
Why it matters
Every published figure for machine reading of cuneiform comes from texts that scholars had already edited: BLEU 36 to 37 for Akkadian royal inscriptions, top-five sign accuracy of 0.80 on the Persepolis tablets, about 1,500 fragment joins proposed by matching and confirmed by hand[1][2][3]. No test has yet run a pipeline from photograph to translation on tablets nobody had read and compared the output with a later expert reading, so whether the backlog can be read by machine is unmeasured[4].
Answers 5
Next steps 5
1
Genre-transfer test: train a transliteration-to-English model on the Akkademia ORACC splits (Neo-Assyrian royal inscriptions), evaluate unchanged on ORACC Old Babylonian letters and on SumTablets-linked Ur III texts with English, and report BLEU/chrF against the in-genre held-out score.
~16 h
—not judged
2
Run the genre-transfer test of Akkadian NMT on Old Babylonian letters and Ur III texts.
~16 h
—not judged
3
Glyph-to-transliteration baseline on the stored SumTablets test split (sha256 65dfe606…): train on the public train split, report sign-reading accuracy by period and genre, and compare with Akkademia's 96.7 per cent on RINAP to see how much reading Sumerian signs costs over Akkadian.
~10 h
—not judged
4
Measure the backlog: from the CDLI bulk dump (cdli_cat.csv plus cdliatf_unblocked.atf) tabulate catalogued artifacts by period and genre with the share that has ATF, and reconcile the Ur III row with BDTNS's 105,187 / 75,818 figures.
~6 h
—not judged
5
Design a blind held-out benchmark: ask a holding museum (British Museum or Penn) for photographs of 200 never-transliterated tablets, run sign detection plus transliteration, and have two specialists transliterate the same tablets independently to score agreement.
~40 h
—not judged
Ranked by how likely each is to move the key question, for the time it takes
History
- Manuel Molina: the Ur III corpus, the largest unread block, is at least 120,000 administrative tablets of which 75,818 are…For: The backlog needs context, not translation
- Ravneet Punia et al.: a first Sumerian machine-translation system for Ur III administrative transliterations produced translations its…For: Machine translation scales to the backlog
- Ethan Fetaya et al.: both restoration studies (recurrent networks 2020, masked language models 2021) present their output as suggestions…For: Machines as assistants, reading stays human
- Koren Lazar et al.: both restoration studies (recurrent networks 2020, masked language models 2021) present their output as suggestions…For: Machines as assistants, reading stays human
- Gai Gutherz et al.: the same study reports hallucinated output, weaker results on long and logogram-heavy lines, and test sentences drawn…Against: Machine translation scales to the backlog
- Gai Gutherz et al.: neural translation of Akkadian from five ORACC corpora reached BLEU 37.47 (transliteration) and 36.52 (cuneiform) on…For: Machine translation scales to the backlog
- Edward C. Williams et al.: deepScribe's chained pipeline on the Persepolis Fortification tablets fell from classifier top-five accuracy 0.89 to…For: The bottleneck is sign detection, not translation
- Ludwig-Maximilians-Universität München: the eBL project reports about 1,500 fragment joins in five years against about 5,000 in the previous 150 years, each…For: Machines as assistants, reading stays human
- Electronic Babylonian Library project team: the eBL project reports about 1,500 fragment joins in five years against about 5,000 in the previous 150 years, each…For: Machines as assistants, reading stays human
- Fabian Simonjetz et al.: the eBL Fragmentarium's n-gram matcher surfaced three unidentified Maqlû manuscripts when an expert reviewed the top…For: Machines as assistants, reading stays human
- Manuel Molina: the Ur III corpus, the largest unread block, is at least 120,000 administrative tablets of which 75,818 are…For: The backlog needs context, not translation
- Wentao Che et al.: the 2026 eBL-wide detector reports relative gains over prior work but no absolute sign accuracy, and its authors note…For: The bottleneck is sign detection, not translation
- Session 1: raised the question and proposed five answers