# The unread cuneiform: half a million tablets and the limits of Sumerian

> Collections hold more than 500,000 cuneiform tablets, written from about 3300 BC to the first century AD, and most have never been published or…

Source: https://mysteriesdept.com/mysteries/unread-cuneiform  
Publisher: Mysteries Department  
Licence: CC BY 4.0 (credit Mysteries Department and link the page)  
Updated: 2026-10-31T16:16:54+00:00  

---

**MYS** 0DEDMM05PWD2·Mystery · document

# The unread cuneiform: half a million tablets and the limits of Sumerian

**0** [Follow][Ask a question](https://mysteriesdept.com/questions/key-questions?mystery=unread-cuneiform)

The unread cuneiform: half a million tablets and the limits of Sumerian

![Photograph of an Ur III Sumerian administrative tablet, c. 2100 to 2000 BC](https://mysteriesdept.com/media/folio/unread-cuneiform-lead)

SpeakingArch · CC BY-SA 4.0

Where

Iraq, Syria, Iran, Turkey and the Levant; held mainly in London, New Haven, Philadelphia, Berlin, Paris, Istanbul and Baghdad⁠[[5]]

When

c. 3300 BC to the 1st century AD⁠[[13]]

Corpus

More than 500,000 (CDLI) or 550,000 (York, 2026) inscribed objects in collections⁠[[4]]

Catalogued

More than 400,000 on CDLI, with about 2 million lines of transliteration⁠[[1]]

Ur III administrative texts

At least 120,000 tablets; 105,187 catalogued and 75,818 transliterated in BDTNS (May 2026)⁠[[6]]

Sumerian literature edited

About 400 compositions, roughly 50,000 lines (ETCSL, 1997 to 2006)⁠[[21]]

Sumerian lexicon

About 16,000 words and 3.4 million word instances from 110,000 manuscripts (ePSD2 2.7.2)⁠[[3]]

Fragmentarium

About 27,000 transliterated fragments of first-millennium Akkadian texts (eBL, 2024)⁠[[12]]

Languages

Sumerian (isolate), Akkadian (Semitic), Elamite, Hittite, Hurrian, Urartian and others⁠[[16]]

Hardest texts

Early Dynastic literary tablets from Fara and Abu Salabikh, c. 2600 BC, in UD.GAL.NUN spelling⁠[[8]]

Status

Unexplained

Questions

0 of 7 answered

People [**Bram Jagersma** Leiden Sumerologist whose 2010 descriptive grammar is the fullest modern account of the language.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNGNN77HF/bram-jagersma)[**Manfred Krebernik** Jena Assyriologist and the leading editor of Early Dynastic literary and incantation texts from Fara, Ebla and Abu Salabikh.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNG533X1W/manfred-krebernik)[**Robert D. Biggs** Oriental Institute Assyriologist who published the Abu Salabikh tablets and the archaic Kesh Temple Hymn.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNFSKFEYQ/robert-d-biggs)[**Enrique Jiménez** Assyriologist at LMU Munich who leads the electronic Babylonian Library and its Fragmentarium.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNESA3K4W/enrique-jimenez)[**Manuel Molina** CSIC Assyriologist who built and edits the Database of Neo-Sumerian Texts, the census of the Ur III corpus.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNH1AZ598/manuel-molina)

All 7 people ›

[**Robert K. Englund** Assyriologist at UCLA, founding director of the Cuneiform Digital Library Initiative and editor of the archaic Uruk texts.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNFAB6CHW/robert-k-englund)[**Shai Gordin** Assyriologist at Ariel University who leads the Digital Pasts Lab and co-authored the neural restoration and Akkadian machine-translation papers.](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNF24S7H7/shai-gordin)

Places [All 7](https://mysteriesdept.com/places?mystery=unread-cuneiform) [**Tell Abu Salabikh, Iraq** site](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDMZXFEW7W/tell-abu-salabikh-iraq)[**British Museum, London** museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN19B9HCV/british-museum-london)[**Iraq Museum, Baghdad** museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN0X5DGC2/iraq-museum-baghdad)[**Penn Museum, Philadelphia** museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN1FXGMD1/penn-museum-philadelphia)

Organisations [**Cuneiform Digital Library Initiative** ](https://mysteriesdept.com/records/ORG0D1P8Z1E15FH/cuneiform-digital-library-initiative)

Concepts [All](https://mysteriesdept.com/concepts)

[Held-out test](https://mysteriesdept.com/concepts/CPT0DECW06RF522/held-out-test)[Language isolate](https://mysteriesdept.com/concepts/CPT0DEEBNN2HBN9/language-isolate)[Null model](https://mysteriesdept.com/concepts/CPT0DECW06JED24/null-model)[UD.GAL.NUN orthography](https://mysteriesdept.com/concepts/CPT0DEDZ5HH5Q47/ud-gal-nun-orthography)

Catalog key**MYS** 0DEDMM05PWD2

Cuneiform is the earliest writing system that can be read in bulk. It was deciphered in the middle of the nineteenth century, its two main languages, Sumerian and Akkadian, have dictionaries and grammars, and three online corpora (CDLI, ORACC and ePSD2) give open transliterations of hundreds of thousands of texts⁠[[1]]⁠[[2]]⁠[[3]]. What has not happened is the reading of the corpus itself. CDLI estimates that collections hold more than 500,000 inscribed objects; it has catalogued more than 400,000 and holds about two million lines of transliteration, which is a fraction of what the tablets carry⁠[[1]]⁠[[4]]. The [British Museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN19B9HCV/british-museum-london) alone holds about 130,000 tablets, many of them unstudied since they arrived in the 1870s and 1880s⁠[[5]].

The backlog is uneven. The largest block is the administrative output of the Ur III state (about 2100 to 2000 BC): at least 120,000 tablets, of which the BDTNS database had transliterated 75,818 by May 2026⁠[[6]]⁠[[7]]. These are short, formulaic and readable, and the obstacle is labour, not knowledge. At the other extreme are the Early Dynastic literary tablets from Fara and Abu Salabikh (about 2600 BC), which were recognised as literature only in the 1960s and 1970s and which are written in a spelling, called [UD.GAL.NUN](https://mysteriesdept.com/concepts/CPT0DEDZ5HH5Q47/ud-gal-nun-orthography) after three of its substituted signs, that omits grammatical elements and does not keep the signs in spoken order; their modern editors describe their translations as provisional⁠[[8]]⁠[[9]]⁠[[10]]. Between the two lie tens of thousands of broken first-millennium tablets of Akkadian literature and scholarship, whose text can only be recovered when fragments scattered across museum drawers are joined⁠[[11]]⁠[[12]].

Beneath the tablets lies the language problem. Sumerian has no known relative; every affiliation proposed since the nineteenth century has been rejected, and the idea of a pre-Sumerian substratum language in southern Iraq did not survive Rubio's 1999 review of the evidence⁠[[13]]⁠[[14]]⁠[[15]]. Its phonology is inferred from how Akkadian scribes transcribed it and from spelling variation, so several consonants and the vowel system remain uncertain, and the verbal prefix chain is analysed differently in each of the three main grammars⁠[[16]]⁠[[13]]. Akkadian is better understood, but its signs are polyvalent and its texts are broken, so that a translation of a damaged line is always an interpretation⁠[[17]].

Since 2018 the question has been whether machines can take up the reading. The eBL Fragmentarium in Munich matched about 1,500 fragments to known texts in five years, against about 5,000 joins in the previous 150 years of Assyriology⁠[[18]]. Neural translation of Neo-Assyrian royal inscriptions reaches BLEU scores in the high thirties on [held-out](https://mysteriesdept.com/concepts/CPT0DECW06RF522/held-out-test) sentences, sign detectors locate and classify signs on photographs with top-five accuracy around 80 per cent, and language models propose fillings for gaps⁠[[17]]⁠[[19]]⁠[[20]]. None of these yet reads an unpublished tablet end to end, and all were trained on the genres already best edited. Whether they can be extended to the unpublished majority, and whether any method can read the Early Dynastic texts whose spelling omits what grammar needs, are the open questions. Of the 7 questions on this site, 0 are answered.

### Key questionsset by the editors

[**Can machines read the cuneiform tablets that no scholar has read?** ](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2BJHV5Z/can-machines-read-the-cuneiform-tablets-that-no-scholar-has)No answer ahead yet

[**Can the broken tablets of Babylonian literature be reassembled by automated matching of fragments?** ](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3VBZYJY/can-the-broken-tablets-of-babylonian-literature-be)No answer ahead yet

[**How many cuneiform tablets survive, and what share has been published or transliterated?** ](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2T64VQB/how-many-cuneiform-tablets-survive-and-what-share-has-been)No answer ahead yet

[**What about the Sumerian language itself is still unknown after 170 years of study?** ](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3N5AH6P/what-about-the-sumerian-language-itself-is-still-unknown)No answer ahead yet

[**Why do the Early Dynastic literary texts from Fara and Abu Salabikh resist reading, and can they be read without later duplicates?** ](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3AM4NHK/why-do-the-early-dynastic-literary-texts-from-fara-and-abu)No answer ahead yet

 [All 7 questions ](https://mysteriesdept.com/questions/key-questions?mystery=unread-cuneiform)

### Research sessions1 so far · last 31 Oct

`S-001`[The unread cuneiform seeded: corpus, readings and open data](https://mysteriesdept.com/mysteries/unread-cuneiform/sessions/SES0DEE9K0P47XS/session-1-the-unread-cuneiform-seeded-corpus-readings-and)

31 Oct**7** questions raised**12** answers proposedNo question changed standing

 [All 1 sessions ](https://mysteriesdept.com/mysteries/unread-cuneiform/sessions)

## The object

The corpus is every clay tablet, prism, brick, seal and stone inscribed with wedge-shaped signs between about 3300 BC and the first century AD. No official count exists. CDLI estimates that public and private collections hold more than 500,000 exemplars, and the 2026 York programme puts the figure above 550,000; CDLI has catalogued more than 400,000 and holds about two million lines of transliteration⁠[[1]]⁠[[4]]. The British Museum holds about 130,000 tablets and the Yale Babylonian Collection about 40,000, followed by the [Penn Museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN1FXGMD1/penn-museum-philadelphia), the Vorderasiatisches Museum in Berlin, the Louvre, Istanbul and the [Iraq Museum](https://mysteriesdept.com/mysteries/unread-cuneiform/places/PLC0DEDN0X5DGC2/iraq-museum-baghdad)⁠[[5]]. The largest single block is administrative. The Ur III state (about 2100 to 2000 BC) left at least 120,000 tablets, of which the BDTNS database had catalogued 105,187 and transliterated 75,818 by May 2026⁠[[6]]⁠[[7]]. The literary corpus is small by comparison. ETCSL edited about 400 Sumerian compositions, roughly 50,000 lines, most of them known from Old Babylonian school copies of about 1800 BC, before its funding ended in 2006⁠[[21]]. ePSD2, the lemmatised Sumerian corpus behind the Pennsylvania dictionary, now draws on more than 110,000 manuscripts and 3.4 million word instances⁠[[3]]. For Akkadian literature and scholarship of the first millennium BC, the eBL Fragmentarium in Munich held about 27,000 transliterated fragments by 2024, most from the British Museum's Kuyunjik (Nineveh) and Babylon collections⁠[[12]]⁠[[11]]. The oldest literary tablets are about 500 Early Dynastic IIIa pieces from Abu Salabikh, excavated in 1963 and 1965, and a comparable group from Fara (ancient Shuruppak) dug in 1902 and 1903, now in Baghdad, Berlin and Philadelphia⁠[[22]]⁠[[23]]. Transliterations on CDLI may be reused with attribution and ORACC data is CC BY-SA; the photographs remain the museums' property, which limits what image-based work can be published⁠[[24]]⁠[[2]].

### Timeline

**Timeline**

Squeezed where little happened. Each mark is a published claim, by how it stands:AcceptedMost likelyStill possibleDisputedRuled out

## What is known

Cuneiform is read. Akkadian, a Semitic language with living relatives, was deciphered in the 1850s, and Sumerian was recovered through the bilingual word lists and translations that Babylonian scribes themselves compiled; a specialist reads an ordinary account or royal inscription at sight. The uncertainty is at the edges. Sumerian is a [language isolate](https://mysteriesdept.com/concepts/CPT0DEEBNN2HBN9/language-isolate). Proposals to connect it with Indo-European, Dravidian, Uralic, Turkic, Austroasiatic, Sino-Tibetan and various macrofamilies have all been rejected, and Whittaker's suggestion of an early Indo-European language ("Euphratic") behind the Uruk texts has no following⁠[[13]]⁠[[15]]. Landsberger's pre-Sumerian substratum, meant to explain polysyllabic words and old city names, was taken apart item by item by Rubio in 1999⁠[[14]]. Sumerian phonology is reconstructed from Akkadian transcriptions and spelling variation, so the consonant transcribed ř, the velar nasal, the number of vowels and the existence of /h/ and a glottal stop are all disputed, and stress is unknown⁠[[13]]⁠[[16]]. The verbal prefix chain is analysed differently in the main grammars of Thomsen (1984), Edzard (2003) and [Jagersma](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNGNN77HF/bram-jagersma) (2010)⁠[[13]]. Emesal, the dialect of laments and of goddesses' speech, was spelt out syllable by syllable because ordinary word signs would be read in the main dialect by default, so it is well attested but its relation to the main dialect is still argued⁠[[13]]⁠[[25]]. Deimel counted about 870 signs for Early Dynastic IIIa⁠[[13]]. The lexical lists, the scribes' own dictionaries, run from the archaic Uruk lists to first-millennium commentaries; the profession list Lú A was copied almost unchanged for a thousand years, and at least one archaic list, AD-GI4, may not be Sumerian at all⁠[[26]]. Whether proto-cuneiform encodes a language is unsettled⁠[[13]].

## What has been tested

Two kinds of test have been run: on the hardest texts, and on the tools meant to speed up reading of the rest. On the Early Dynastic literary tablets, Krecher showed in 1978 that the Fara and Abu Salabikh compositions use a regular system of sign substitutions, UD for AN, GAL for EN, NUN for LIL and others, which has to be undone before the text can be read⁠[[8]]. Alster asked in 1982 whether the system hid an early form of Emesal; the question is still open⁠[[25]]. [Biggs](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNFSKFEYQ/robert-d-biggs) showed in 1971 that the Abu Salabikh Kesh Temple Hymn matches the Old Babylonian text eight centuries later except in spelling, which is why that one Early Dynastic composition can be read⁠[[27]]. [Krebernik](https://mysteriesdept.com/mysteries/unread-cuneiform/people/FIG0DEDNG533X1W/manfred-krebernik) and Lisman's 2020 edition of the 70 Zame hymns, which have no later parallel, calls its translation not definitive because the orthography omits suffixes, postpositions and verbal morphemes and does not keep the signs in spoken order⁠[[9]]⁠[[23]]. Cooper's judgement is that the UD.GAL.NUN texts remain very hard to understand and probably went out of use before Ur III⁠[[10]].

On the tools, Fetaya and colleagues trained recurrent networks on Achaemenid Babylonian archival texts in 2020 to propose fillings for breaks, and Lazar and colleagues did the same with masked language models in 2021⁠[[20]]⁠[[28]]. Punia and colleagues built the first Sumerian machine translation, for Ur III administrative texts, in 2020⁠[[29]]. Gutherz and colleagues trained neural translation on 56,160 Akkadian sentences from five ORACC corpora and reached BLEU 37.47 from transliteration and 36.52 from cuneiform signs, against a translation-memory [baseline](https://mysteriesdept.com/concepts/CPT0DECW06JED24/null-model) of 23.51 and 27.09; they report hallucinated output and poorer results on long or logogram-heavy lines⁠[[17]]⁠[[30]]. DeepScribe, trained on 100,000 annotated signs from the Persepolis Fortification tablets, located signs with mAP 0.78 and classified them with top-five accuracy 0.89, falling to 0.80 end to end⁠[[19]]. The eBL Fragmentarium's matching, a bag of character n-grams scored by the overlap coefficient, surfaced three unidentified manuscripts of the Maqlû incantations when an expert checked the top 300 of some 25,000 candidates, and the project reports about 1,500 joins in five years⁠[[12]]⁠[[18]]. In 2026 a sign detector was run over 87,668 eBL fragments, giving 2.9 million detections, with the authors noting its sensitivity to damage⁠[[31]]. Heidelberg's HeiCuBeDa provides 3D scans and six-view images of Hilprecht Collection tablets as a benchmark⁠[[32]].

## The readings

On whether machines can read the unread tablets, four positions are live. The first holds that neural translation already produces useful drafts for well-edited genres and will scale to the backlog as training data grows; it rests on the Akkadian results and on the fact that most of the backlog is formulaic administrative text⁠[[17]]⁠[[29]]. The second holds that reading is a pipeline from photograph to signs to transliteration to translation, and that the first stage, sign detection on damaged, curved, variably lit clay, is the bottleneck; DeepScribe's drop from 0.89 to 0.80 when its stages are chained, and the 2026 detector's sensitivity to damage, are the evidence⁠[[19]]⁠[[31]]. The third holds that the realistic role of machines is assistance: proposing joins, fillings and sign readings for a human to check, which is how the Fragmentarium's 1,500 joins were made and how the restoration models were evaluated⁠[[12]]⁠[[20]]⁠[[28]]. The fourth holds that the backlog is mostly administrative text whose difficulty is not translation but interpretation, which needs names, year dates and archive context that no model has, so that machines will catalogue and transliterate while the reading stays human⁠[[6]]⁠[[7]].

On the Early Dynastic literary texts, three positions. One treats UD.GAL.NUN as a learned orthography, a set of substitutions that can in principle be fully tabulated, after which the texts can be read like any Sumerian⁠[[8]]. A second takes it as evidence of a dialect, perhaps early Emesal, which would make the texts a different language variety from the later copies⁠[[25]]. A third holds that, whatever the system, texts without a later parallel cannot be read with confidence because the spelling omits what grammar needs, so only compositions like the Kesh hymn, with Old Babylonian duplicates, will ever be secure⁠[[27]]⁠[[9]]⁠[[10]].

On Sumerian's origins, the default is that it is an isolate with no detectable substratum⁠[[16]]⁠[[14]]; Whittaker's Euphratic is the only live minority view⁠[[15]].

### Published claims

**Published claims and their status**

| Year | Who | Claimed | Status | Reason |
|---|---|---|---|---|
| 1971 | Robert D. Biggs⁠[27] | The Abu Salabikh copies of the Kesh Temple Hymn match the Old Babylonian text of eight centuries later apart from spelling, so this one Early Dynastic composition can be read. | Accepted | The identification has held, and the CDLI witnesses are transliterated against the composite⁠[36]. |
| 1978 | Joachim Krecher⁠[8] | The Fara and Abu Salabikh literary texts use a regular system of sign substitutions (UD.GAL.NUN) that must be undone before reading. | Accepted | Every later edition of these texts works from Krecher's system, though the table is still incomplete⁠[9]. |
| 1982 | Bendt Alster⁠[25] | UD.GAL.NUN spelling may encode an early form of the Emesal dialect. | Disputed | The idea has not been confirmed or refuted; the texts omit too much grammar to test dialect features⁠[23]. |
| 1999 | Gonzalo Rubio⁠[14] | The words cited for a pre-Sumerian substratum are loans, compounds or otherwise explicable, so no substratum is demonstrated. | Most likely | Most Sumerologists accept it; Whittaker objects that the argument from phonotactics is circular⁠[15]. |
| 2005 | Gordon Whittaker⁠[15] | About 200 early loanwords show an Indo-European language (Euphratic) in contact with Sumerian in the Uruk period. | Disputed | Mainstream opinion rejects it, chiefly because the dates are too early for Indo-European and the matches are unsystematic⁠[13]. |
| 2008 | Manuel Molina⁠[7] | The Ur III corpus runs to well over 100,000 tablets, most of them administrative, distributed across a handful of provincial archives. | Accepted | BDTNS now catalogues 105,187 and estimates at least 120,000 in collections⁠[6]. |
| 2010 | Bram Jagersma⁠[16] | Sumerian is an isolate, a remnant of the languages spoken in southern Iraq before Semitic arrived, with a phoneme inventory that can be reconstructed from spelling and transcription. | Accepted | It is the fullest grammar and its isolate verdict is the consensus, but several of its phonemes are contested⁠[13]. |
| 2020 | Ethan Fetaya, Yonatan Lifshitz, Elad Aaron and Shai Gordin⁠[20] | Recurrent networks trained on Achaemenid Babylonian texts can propose plausible fillings for broken passages. | Still possible | Evaluated on artificially masked text, not on tablets later found whole⁠[28]. |
| 2020 | Ravneet Punia, Niko Schenk, Christian Chiarcos and Émilie Pagé-Perron⁠[29] | Ur III administrative Sumerian can be machine-translated into usable English. | Still possible | The texts are formulaic and the test set is drawn from the same archives; transfer to other genres is untested⁠[29]. |
| 2021 | Koren Lazar and colleagues⁠[28] | Masked language models fine-tuned on Akkadian transliterations propose correct words for gaps often enough to help editors. | Still possible | Judged by held-out masking and a small expert evaluation; no published edition yet credits a model-proposed restoration⁠[28]. |
| 2023 | Gai Gutherz, Shai Gordin, Luis Sáenz, Omer Levy and Jonathan Berant⁠[17] | Neural machine translation of Akkadian royal inscriptions and letters reaches BLEU 36 to 37 and beats a translation-memory baseline. | Most likely | The scores are on held-out sentences from the same five corpora, and the authors report hallucinations and weakness on long lines⁠[17]. |
| 2023 | Edward Williams and colleagues (DeepScribe)⁠[19] | Signs on photographs of Persepolis Fortification tablets can be located (mAP 0.78) and classified (top-5 0.89), with 0.80 end to end. | Accepted | The archive is unusually homogeneous in hand, period and language, so the result may not carry to mixed collections⁠[31]. |
| 2023 | Noah Hermalin⁠[35] | The degree of logography of Sumerian writing can be measured by mutual information between signs and the words they write. | Still possible | A method paper on two languages; not yet applied across periods of Sumerian⁠[35]. |
| 2024 | Fabian Simonjetz, Jussi Laasonen, Yunus Cobanoglu, Alexander Fraser and Enrique Jiménez⁠[12] | Character n-gram overlap matching over the Fragmentarium finds unidentified manuscripts fast enough to change how Assyriologists work. | Accepted | Three new Maqlû manuscripts came from one expert pass over 300 candidates, and the project counts about 1,500 joins⁠[18]. |
| 2026 | Wentao Che, Esteban Garcés Arias, Asim Niaz, Andreas Bender and Enrique Jiménez⁠[31] | A Deformable DETR detector can be run over 87,668 eBL fragments to produce 2.9 million sign detections. | Still possible | A preprint without peer review; the authors note sensitivity to damage and layout and give no transliteration accuracy⁠[31]. |
| 2020 | Manfred Krebernik and Jan Lisman⁠[9] | The 70 Zame hymns from Abu Salabikh can be edited and translated, but only provisionally, because their orthography omits grammar and reorders signs. | Accepted | The editors say so themselves, as Biggs did of his 1974 first edition⁠[23]. |

## What would settle it

For the machine question, a held-out test is the obvious instrument: take tablets that have been photographed but never transliterated, run the full pipeline, then have specialists transliterate the same tablets blind and score the agreement; no such test has been published, and every reported figure comes from texts already edited⁠[[17]]⁠[[31]]⁠[[19]]. A second test is genre transfer: models trained on Neo-Assyrian royal inscriptions or Persepolis Elamite applied to Old Babylonian letters or Ur III accounts, with the error rate reported. The data for both exist. CDLI's bulk dump gives every public transliteration, SumTablets gives 91,606 Sumerian tablets as glyph-transliteration pairs under CC BY 4.0, and the Akkademia repository gives the ORACC training sets⁠[[33]]⁠[[34]]⁠[[30]]. What is missing is a benchmark of unpublished tablets with later expert readings, which only a museum partnership can provide.

For the Early Dynastic texts, the decisive evidence would be new duplicates: an Old Babylonian copy of a Zame hymn or of a UD.GAL.NUN myth would do for those texts what the Old Babylonian Kesh hymn did for its Abu Salabikh ancestor⁠[[27]]. Short of that, a complete table of UD.GAL.NUN substitutions tested against the whole Fara and Abu Salabikh corpus would show whether the system is regular enough to be an orthography or variable enough to be something else⁠[[8]]⁠[[25]].

For the Sumerian question, nothing short of a related language could settle affiliation, and none is expected. What can be settled is phonology, by systematic study of how Akkadian and other scribes transcribed Sumerian, and by the kind of information-theoretic measurement of the script that Hermalin applied to Sumerian in 2023⁠[[35]]⁠[[16]].

## Questions

0 of 7 questions are answered.

**Key questions** [All 7](https://mysteriesdept.com/questions/key-questions?mystery=unread-cuneiform)

 **0**

[Can machines read the cuneiform tablets that no scholar has read?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2BJHV5Z/can-machines-read-the-cuneiform-tablets-that-no-scholar-has)

0 of 0 sub-questions answered·0

Open: 5 answers, none ahead



 **0**

[Can the broken tablets of Babylonian literature be reassembled by automated matching of fragments?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3VBZYJY/can-the-broken-tablets-of-babylonian-literature-be)

0 of 0 sub-questions answered·0

Open: 2 answers, none ahead



 **0**

[How many cuneiform tablets survive, and what share has been published or transliterated?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2T64VQB/how-many-cuneiform-tablets-survive-and-what-share-has-been)

0 of 0 sub-questions answered·0

Open: no answer ahead yet



 **0**

[What about the Sumerian language itself is still unknown after 170 years of study?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3N5AH6P/what-about-the-sumerian-language-itself-is-still-unknown)

0 of 2 sub-questions answered·0

Open: no answer ahead yet



 **0**

[Why do the Early Dynastic literary texts from Fara and Abu Salabikh resist reading, and can they be read without later duplicates?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3AM4NHK/why-do-the-early-dynastic-literary-texts-from-fara-and-abu)

0 of 0 sub-questions answered·0

Open: 3 answers, none ahead



## Evidence 23 findings

[Can machines read the cuneiform tablets that no scholar has read?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2BJHV5Z/can-machines-read-the-cuneiform-tablets-that-no-scholar-has)10 findings

[Machine translation scales to the backlog](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDW61E16Z8/machine-translation-scales-to-the-backlog)Still possible

+[Neural translation of Akkadian from five ORACC corpora reached BLEU 37.47 (transliteration) and 36.52 (cuneiform) on held-out sentences, against 23.51 and 27.09 for a translation-memory baseline.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXNRE4WCG/neural-translation-of-akkadian-from-five-oracc-corpora)Supported

+[A first Sumerian machine-translation system for Ur III administrative transliterations produced translations its authors judged useful in quantitative and qualitative evaluation.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXP3V2WH7/a-first-sumerian-machine-translation-system-for-ur-iii)Suggestive

−[The same study reports hallucinated output, weaker results on long and logogram-heavy lines, and test sentences drawn from the training corpora, so no cross-genre or unpublished-text performance is known.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXNXMGRAR/the-same-study-reports-hallucinated-output-weaker-results-on)Supported

[The bottleneck is sign detection, not translation](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDW6Q19ZZH/the-bottleneck-is-sign-detection-not-translation)Still possible

+[DeepScribe's chained pipeline on the Persepolis Fortification tablets fell from classifier top-five accuracy 0.89 to 0.80 end to end, with no top-one figure reported, on an archive uniform in hand, period and language.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXP94PC2X/deepscribes-chained-pipeline-on-the-persepolis-fortification)Supported

+[The 2026 eBL-wide detector reports relative gains over prior work but no absolute sign accuracy, and its authors note sensitivity to tablet damage and layout.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXPHRKAXK/the-2026-ebl-wide-detector-reports-relative-gains-over-prior)Suggestive

+[Transliteration from already-digitised Unicode signs reaches 96.7 per cent with a BiLSTM, far above any reported photograph-based sign accuracy, so the gap lies before the transliteration stage.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXPP57XVA/transliteration-from-already-digitised-unicode-signs-reaches)Supported

[Machines as assistants, reading stays human](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDW6Z7X8W5/machines-as-assistants-reading-stays-human)Still possible

+[The eBL Fragmentarium's n-gram matcher surfaced three unidentified Maqlû manuscripts when an expert reviewed the top 300 of about 25,000 candidates; the identifications were made by the expert, not the matcher.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDY1GG73K2/the-ebl-fragmentariums-n-gram-matcher-surfaced-three)Supported

+[The eBL project reports about 1,500 fragment joins in five years against about 5,000 in the previous 150 years, each confirmed by a specialist.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDY2047S74/the-ebl-project-reports-about-1-500-fragment-joins-in-five)Suggestive

+[Both restoration studies (recurrent networks 2020, masked language models 2021) present their output as suggestions for editors and evaluate them by masked held-out text and expert judgement, not by adoption in editions.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDY3FKDT9Y/both-restoration-studies-recurrent-networks-2020-masked)Suggestive

[The backlog needs context, not translation](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDW7V5CFDG/the-backlog-needs-context-not-translation)Still possible

+[The Ur III corpus, the largest unread block, is at least 120,000 administrative tablets of which 75,818 are transliterated and 105,187 catalogued; its texts are short, formulaic and lexically known.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDY48E588Q/the-ur-iii-corpus-the-largest-unread-block-is-at-least-120)Supported

[How many cuneiform tablets survive, and what share has been published or transliterated?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT2T64VQB/how-many-cuneiform-tablets-survive-and-what-share-has-been)1 finding

[500,000+ survive; under half transliterated](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEEANK5JS75/500-000-survive-under-half-transliterated)Still possible

+[CDLI estimates more than 500,000 exemplars and has catalogued more than 400,000; York's 2026 figure is above 550,000; BDTNS transliterates 75,818 of 105,187 catalogued Ur III texts.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEEB0X4E2DT/cdli-estimates-more-than-500-000-exemplars-and-has)Supported

[Why do the Early Dynastic literary texts from Fara and Abu Salabikh resist reading, and can they be read without later duplicates?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3AM4NHK/why-do-the-early-dynastic-literary-texts-from-fara-and-abu)4 findings

[UD.GAL.NUN is a recoverable orthography](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDWTEHR4KW/ud-gal-nun-is-a-recoverable-orthography)Still possible

+[Krecher identified regular sign substitutions (UD for AN, GAL for EN, NUN for LIL) across the Fara and Abu Salabikh literary tablets, and every later edition works from his system.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYQRHMP6Q/krecher-identified-regular-sign-substitutions-ud-for-an-gal)Supported

−[Even where the substitutions are undone, the Early Dynastic spelling omits suffixes, postpositions and verbal morphemes and does not keep sign order, so the 2020 editors of the Zame hymns still call their translation provisional.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYR66DXTW/even-where-the-substitutions-are-undone-the-early-dynastic)Supported

[Only texts with later duplicates can be read](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDWX41X1XH/only-texts-with-later-duplicates-can-be-read)Still possible

+[The Abu Salabikh Kesh Temple Hymn is readable because its Old Babylonian duplicate of eight centuries later matches it except in spelling; Biggs said these early texts present formidable obstacles to comprehension.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYS8MZGTA/the-abu-salabikh-kesh-temple-hymn-is-readable-because-its)Supported

+[Cooper judges the UD.GAL.NUN compositions very hard to understand and thinks they left the curriculum before Ur III, which would mean no later duplicates exist to be found.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYSNXZZ9Q/cooper-judges-the-ud-gal-nun-compositions-very-hard-to)Suggestive

[What about the Sumerian language itself is still unknown after 170 years of study?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3N5AH6P/what-about-the-sumerian-language-itself-is-still-unknown)1 finding

[Words known; sounds, verb and origin disputed](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEEAR5X6GAT/words-known-sounds-verb-and-origin-disputed)Still possible

+[ePSD2 lists about 16,000 Sumerian words in 3.4 million instances, while several phonemes and the verbal prefix system are analysed differently in the main grammars.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEEB27SHD8S/epsd2-lists-about-16-000-sumerian-words-in-3-4-million)Supported

[Can the broken tablets of Babylonian literature be reassembled by automated matching of fragments?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT3VBZYJY/can-the-broken-tablets-of-babylonian-literature-be)2 findings

[Matching works, but only on transliterated fragments](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEEAA05JR5T/matching-works-but-only-on-transliterated-fragments)Still possible

+[The matcher works on transliterations: three Maqlû manuscripts were found among about 27,000 transliterated fragments, and the project reports about 1,500 joins in five years.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEEAXQEF6RJ/the-matcher-works-on-transliterations-three-maqlu)Supported

[Image-fed matching reaches the untransliterated fragments](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEEAFRT3SF6/image-fed-matching-reaches-the-untransliterated-fragments)Still possible

+[A 2026 pipeline links detected signs to text structure by n-gram similarity across 87,668 eBL fragments, the precondition for image-fed matching, but reports no joins found this way.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEEAZT9E1VT/a-2026-pipeline-links-detected-signs-to-text-structure-by-n)Speculative

[Does Sumerian have any demonstrable relative or substratum, or is it an isolate?](https://mysteriesdept.com/mysteries/unread-cuneiform/questions/QST0DEDT86CFAB2/does-sumerian-have-any-demonstrable-relative-or-substratum)5 findings

[Sumerian is an isolate](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDWXDKH4P6/sumerian-is-an-isolate)Still possible

+[Rubio's 1999 examination of the words cited for a pre-Sumerian substratum found them to be Semitic loans, Sumerian compounds or otherwise explicable, and is the standard reference against the substratum.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYTNSDG8D/rubios-1999-examination-of-the-words-cited-for-a-pre)Supported

+[Every genetic affiliation proposed for Sumerian (Indo-European, Dravidian, Uralic, Turkic, Austroasiatic, Sino-Tibetan, macrofamilies) has been rejected, and the fullest modern grammar treats the language as an isolate.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYVK2TPHT/every-genetic-affiliation-proposed-for-sumerian-indo)Supported

[Whittaker's Euphratic contact language](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDWYC3NPZ9/whittakers-euphratic-contact-language)Still possible

+[Whittaker presents about 200 proposed loanwords and argues that loans can only be demonstrated by identifying a donor language in which they are segmentable.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYX61W8XG/whittaker-presents-about-200-proposed-loanwords-and-argues)Speculative

−[Mainstream opinion rejects Euphratic, chiefly because Indo-European is not expected in fourth-millennium Iraq and the proposed matches are not systematic; Whittaker himself reports that Rubio rejected it.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYWEH6TNJ/mainstream-opinion-rejects-euphratic-chiefly-because-indo)Supported

[A pre-Sumerian substratum](https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDWZHME0M8/a-pre-sumerian-substratum)Still possible

−[Rubio's item-by-item review left no coherent substrate vocabulary; most Sumerologists treat the substratum as unproven.](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDYYQGASG8/rubios-item-by-item-review-left-no-coherent-substrate)Supported

## Artifacts

The films, photos, papers and objects this mystery rests on, each with its own key and its distance from the original.

**Artifacts** [See all 6](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts)

Photo1 step from object

[The Ur III administrative archives (Umma, Girsu, Puzrish-Dagan, Ur, Nippur)](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts/ART0DEDQR8ZDCWZ/the-ur-iii-administrative-archives-umma-girsu-puzrish-dagan)

SpeakingArch (Wikimedia Commons) · CC BY-SA 4.0

Transcription1 step from object

[The Abu Salabikh literary tablets (IAS), including the Kesh Temple Hymn and the Zame hymns](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts/ART0DEDQRQ9Q3ZW/the-abu-salabikh-literary-tablets-ias-including-the-kesh)

Cuneiform Digital Library Initiative, after Biggs 1971 and 1974 · CDLI terms: transliteration freely reusable with attribution

Document1 step from object

[The Fara (Shuruppak) UD.GAL.NUN literary tablets](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts/ART0DEDQYAH6140/the-fara-shuruppak-ud-gal-nun-literary-tablets)

Vorderasiatisches Museum, Berlin (VAT numbers), with pieces in Istanbul and Philadelphia

Dataset1 step from object

[The Kuyunjik collection fragments (British Museum K., Sm., Rm. and DT numbers)](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts/ART0DEDQZ0V8QWY/the-kuyunjik-collection-fragments-british-museum-k-sm-rm-and)

electronic Babylonian Library, LMU Munich · eBL / LMU Munich; photographs copyright the Trustees of the British Museum; access to some fragments restricted

 [See all 6 artifacts](https://mysteriesdept.com/mysteries/unread-cuneiform/artifacts)

## See also

 [

**Proto-cuneiform: the archaic texts of Uruk** The oldest writing, 3350 to 3000 BC: numbers and goods are read, the language is not.

](https://mysteriesdept.com/mysteries/proto-cuneiform) [

**Hattic, Kassite and the half-read languages of the cuneiform archives** Languages written in a deciphered script, cuneiform, whose texts still cannot be translated: Hattic, Kassite, Gutian and the new language of Kalašma.

](https://mysteriesdept.com/mysteries/hattic-kassite-lost-languages) [

**Hurrian songs of Ugarit** The oldest notated song, from Ugarit c. 1400 BC: every term is identified, and no two scholars play it the same way.

](https://mysteriesdept.com/mysteries/hurrian-songs) [

**Unread oracle bone characters** The earliest Chinese writing is read, but about two thirds of its 4,500 graphs are not.

](https://mysteriesdept.com/mysteries/oracle-bone-undeciphered)

## References

1. 1.[CDLI's own account: collections are estimated to exceed 500,000 cuneiform exemplars; more than 400,000 (an older passage says 360,000) are catalogued electronically; about two million lines of transliteration; founded by Robert Englund (UCLA) and Jürgen Renn (MPIWG) with a first funded phase in 2000-2003. Used for the headline counts.](https://mysteriesdept.com/sources/SRC0DEDHPH2QMG3/about-cdli-cuneiform-digital-library-initiative)cited 6 times here**SRC** 0DEDHPH2QMG3
1. 2.[ORACC Steering Committee 2026. Open Richly Annotated Cuneiform Corpus (ORACC). University of Pennsylvania Museum and partners.](https://mysteriesdept.com/sources/SRC0D464DYA4RFJ/open-richly-annotated-cuneiform-corpus-oracc)cited 2 times here**SRC** 0D464DYA4RFJ
1. 3.[ORACC-hosted Sumerian dictionary and corpus portal, release 2.7.2 of 31 August 2024, CC BY-SA: almost 16,000 words, phrases and names plus over 50,000 administrative-name entries, in more than 225,000 distinct forms, drawn from over 110,000 Sumerian manuscripts and about 3.4 million word instances. Used for the size of the lemmatised Sumerian corpus and as the source of SumTablets.](https://mysteriesdept.com/sources/SRC0DEDHNFXDR34/the-electronic-pennsylvania-sumerian-dictionary-epsd2)cited 3 times here**SRC** 0DEDHNFXDR34
1. 4.[News item of 25 March 2026 announcing a 2026-2029 programme at the Archaeology Data Service, funded by the Meditor Trust and the University of York, joining CDLI, ETCSL, ORACC and CDLI-ACT, with an emphasis on Arabic-speaking users. States that public and private collections hold more than 550,000 exemplars and CDLI has catalogued more than 400,000. Used for the latest published totals.](https://mysteriesdept.com/sources/SRC0DEDHQ2ATWMM/towards-universal-access-to-cuneiform-heritage-tua-ch)cited 3 times here**SRC** 0DEDHQ2ATWMM
1. 5.[Biblical Archaeology Review library sidebar, undated. Puts the total at about 500,000 tablets with no official count; British Museum about 130,000, Yale Babylonian Collection about 40,000, then Penn Museum, Berlin, Louvre, Istanbul and Baghdad; says many tablets have sat a century unstudied. Used for the per-museum holdings; the figures are journalistic and should be checked against the museums.](https://mysteriesdept.com/sources/SRC0DEDGRFY6V0J/cuneiform-tablets-whos-got-what)cited 1 time here**SRC** 0DEDGRFY6V0J
1. 6.[CSIC database of Ur III administrative texts, online since 2002, hosted at CESGA since February 2025. May 2026 figures: 105,187 documents catalogued, 75,818 with transliterations (70,063 published, 5,755 unpublished or auctioned), 1,350,035 lines; the page estimates at least 120,000 Ur III tablets in museums and private collections plus an unknown number in the Iraq Museum. Copyright Manuel Molina; no open licence stated. Used for the size of the Ur III backlog.](https://mysteriesdept.com/sources/SRC0DEDH8V5MW2C/database-of-neo-sumerian-texts-bdtns-about)cited 10 times here**SRC** 0DEDH8V5MW2C
1. 7.[In S. J. Garfinkle and J. C. Johnson (eds), The Growth of an Early State in Mesopotamia: Studies in Ur III Administration, BPOA 5 (Madrid, CSIC), pp. 19-53. The standard census of the Ur III corpus by provenance, archive and collection, and the basis of the estimate that well over 100,000 tablets survive. Used for the structure of the Ur III corpus.](https://mysteriesdept.com/sources/SRC0DEDH974MX1D/the-corpus-of-neo-sumerian-tablets-an-overview)cited 5 times here**SRC** 0DEDH974MX1D
1. 8.[Bibliotheca Orientalis 35, pp. 155-160. Krecher's identification of the UD.GAL.NUN system: a set of regular sign substitutions (UD for AN "god", GAL for EN "lord", NUN for LIL, hence the name) in Early Dynastic literary tablets from Fara and Abu Salabikh, which makes those texts unreadable until the substitutions are undone. Used for the definition of the UD.GAL.NUN orthography.](https://mysteriesdept.com/sources/SRC0DEDJ8E54H39/sumerische-literatur-der-fara-zeit-die-ud-gal-nun-texte-i)cited 6 times here**SRC** 0DEDJ8E54H39
1. 9.[Dubsar 12, Zaphon, Münster, 278 pages, ISBN 9783963270345. New edition of the 70 Early Dynastic IIIa Zame hymns (2 to 12 lines each) from Abu Salabikh, after Biggs's 1974 first edition; the authors call their translation not definitive because the orthography leaves out grammatical morphemes and does not keep sign order. Used for the state of reading Early Dynastic literature.](https://mysteriesdept.com/sources/SRC0DEDJ5N9KSRA/the-sumerian-zame-hymns-from-tell-abu-salabih-with-an)cited 8 times here**SRC** 0DEDJ5N9KSRA
1. 10.[In K. Ryholt and G. Barjamovic (eds), Problems of Canonicity and Identity Formation in Ancient Egypt and Mesopotamia, CNI Publications 43, Museum Tusculanum Press. Says the Early Dynastic literary texts, above all the UD.GAL.NUN compositions, are very hard to understand and probably fell out of use before Ur III, and that the origin of much of the Old Babylonian Sumerian corpus is unknown because spelling and grammar were modernised in transmission. Used for a senior Sumerologist's statement of ](https://mysteriesdept.com/sources/SRC0DEDJ974THN7/sumerian-literature-and-sumerian-identity)cited 3 times here**SRC** 0DEDJ974THN7
1. 11.[LMU news release, 30 January 2023, announcing the public release of the eBL Fragmentarium: about 22,000 fragments processed since 2018, thousands more photographed with the British Museum and Iraq Museum; the algorithm identified a fragment of the latest known Gilgamesh tablet (130 BC) in November 2022 and 15 new fragments of a hymn to Babylon; Jiménez says thousands of fragments remain unidentified. Used for the Fragmentarium's scale and first results.](https://mysteriesdept.com/sources/SRC0DEDGS748FDC/electronic-babylonian-literature-playing-with-the-source-of)cited 6 times here**SRC** 0DEDGS748FDC
1. 12.[LREC-COLING 2024, pp. 13712-13721, CC BY-NC 4.0. Treats fragment identification as text matching over the eBL Fragmentarium (about 27,000 transliterated fragments); a binary bag of character n-grams (n up to 3) with the overlap coefficient performed best; a test on the Maqlû incantation series had an expert review the top 300 of some 25,000 candidates and found three unidentified manuscripts (K.18050, Sm.1215, Sm.1852). Code: github.com/ElectronicBabylonianLiterature/ngrammatcher. Used for how t](https://mysteriesdept.com/sources/SRC0DEDGT58C4F7/reconstruction-of-cuneiform-literary-texts-as-text-matching)cited 7 times here**SRC** 0DEDGT58C4F7
1. 13.[Used only as a map of the specialist literature: the list of rejected affiliation proposals (Indo-European, Dravidian, Uralic, Turkic, Austroasiatic, Sino-Tibetan and macrofamily claims), the disputed phonemes (the consonant transcribed ř, the velar nasal, the vowel inventory, the lost /h/ and glottal stop), the statement that verbal morphology is hotly disputed, Deimel's count of 870 signs for Early Dynastic IIIa, and the dates of the main grammars (Thomsen 1984, Attinger 1993, Edzard 2003, Jag](https://mysteriesdept.com/sources/SRC0DEDJNJFJGDJ/sumerian-language-wikipedia-article)cited 10 times here**SRC** 0DEDJNJFJGDJ
1. 14.[Journal of Cuneiform Studies 51, pp. 1-16. Examines Landsberger's Proto-Euphratic and Proto-Tigridian substratum words one by one and finds most are Semitic loans, Sumerian compounds or otherwise explicable, leaving no coherent substrate vocabulary. The standard reference against a pre-Sumerian substratum. DOI not shown on the journal page.](https://mysteriesdept.com/sources/SRC0DEDHR5BNVDP/on-the-alleged-pre-sumerian-substratum)cited 4 times here**SRC** 0DEDHR5BNVDP
1. 15.[In W. H. van Soldt (ed.), Ethnicity in Ancient Mesopotamia (RAI 48, Leiden 2002), Nederlands Instituut voor het Nabije Oosten, pp. 409-429. Reviews the debate on whether Sumerians were native to southern Iraq or arrived late, the Landsberger (1944) and Salonen (1968) substratum proposals, Rubio's 1999 refutation, Semitic loanwords, and Whittaker's own "Euphratic" Indo-European hypothesis, which the field rejects. Used for the Sumerian question and the author's own minority position.](https://mysteriesdept.com/sources/SRC0DEDHQE9W64H/the-sumerian-question-reviewing-the-issues)cited 12 times here**SRC** 0DEDHQE9W64H
1. 16.[Leiden University doctoral thesis, 4 November 2010, open access in the Leiden repository. The fullest modern descriptive grammar, based on texts of about 2500 to 2000 BC; treats Sumerian as an isolate and a remnant of pre-Semitic languages of the region; source for the reconstructed phoneme inventory, the lost consonants /h/, /ʔ/ and /j/, and the verbal prefix analysis. Used for what is and is not settled in Sumerian grammar.](https://mysteriesdept.com/sources/SRC0DEDHRSSNWMQ/a-descriptive-grammar-of-sumerian)cited 6 times here**SRC** 0DEDHRSSNWMQ
1. 17.[PNAS Nexus 2(5), pgad096. Neural machine translation of Akkadian from transliteration (T2E) and from Unicode cuneiform (C2E) into English, trained on 56,160 sentences from five ORACC corpora (RINAP, RIAo, RIBo, SAAo, Suhu; 8,056 tablets, mostly Neo-Assyrian). BLEU4 37.47 (T2E) and 36.52 (C2E) against a translation-memory baseline of 23.51 and 27.09; best on sentences under about 118 characters; hallucinations noted. Used for the state of machine translation of Akkadian and its limits.](https://mysteriesdept.com/sources/SRC0DEDGBY5JRWS/translating-akkadian-to-english-with-neural-machine)[doi](https://doi.org/10.1093/pnasnexus/pgad096)cited 7 times here**SRC** 0DEDGBY5JRWS
1. 18.[The Conversation, May 2023, written by the eBL team at LMU. Reports about 1,500 joins in the project's five years against about 5,000 in the previous 150 years of Assyriology, a string-alignment algorithm modelled on bioinformatics sequence alignment, and a separate model trained on under 20,000 lines that resolved ambiguous sign readings with 98 per cent success. Used for the join counts and the project's own account of its methods.](https://mysteriesdept.com/sources/SRC0DEDGTTRNMP2/ai-is-helping-us-read-ancient-mesopotamian-literature)cited 6 times here**SRC** 0DEDGTTRNMP2
1. 19.[arXiv 2306.01268 (v2, February 2025). Modular pipeline (RetinaNet localisation, ResNet classification) trained on more than 5,000 annotated images and 100,000 sign bounding boxes from the Persepolis Fortification Archive (Achaemenid Elamite): localisation mAP 0.78, classifier top-5 accuracy 0.89, end-to-end top-5 0.80. Used as the benchmark for photograph-to-sign reading on a homogeneous archive.](https://mysteriesdept.com/sources/SRC0DEDGCG9A89X/deepscribe-localization-and-classification-of-elamite)cited 3 times here**SRC** 0DEDGCG9A89X
1. 20.[PNAS 117(37). Recurrent neural network language models of Achaemenid-period Babylonian (Akkadian) archival texts, used to propose fillings for broken passages; first machine-learning restoration work on cuneiform. Used for the fragment-restoration line of work.](https://mysteriesdept.com/sources/SRC0DEDGV4PQG6J/restoration-of-fragmentary-babylonian-texts-using-recurrent)[doi](https://doi.org/10.1073/pnas.2003794117)cited 4 times here**SRC** 0DEDGV4PQG6J
1. 21.[Oxford project pages (1997-2006, founded by Jeremy Black): composite transliterations and translations of about 400 Sumerian literary compositions, roughly 50,000 lines, mostly Old Babylonian copies; the project closed in 2006 and the site is frozen. Used for the size of the edited Sumerian literary corpus; the content is mirrored in ORACC as epsd2/literary.](https://mysteriesdept.com/sources/SRC0DEDH9RXCRY0/the-electronic-text-corpus-of-sumerian-literature-about-the)cited 3 times here**SRC** 0DEDH9RXCRY0
1. 22.[Iraq 71, pp. 1-32. Publishes the remaining Abu Salabikh pieces (IAS 533-556), maps the distribution of tablets across the site and what it implies for the buildings' function, and indexes the names and words of the administrative texts. Used for the find context of the Early Dynastic literary tablets.](https://mysteriesdept.com/sources/SRC0DEDJ6QF8S1J/the-tablets-from-abu-salabikh-and-their-provenance)[doi](https://doi.org/10.1017/S0021088900000711)cited 5 times here**SRC** 0DEDJ6QF8S1J
1. 23.[Used as a summary of Krebernik and Lisman 2020: 70 hymns of 2 to 12 lines on 24 tablets and fragments (at least seven copies) found in one room at Abu Salabikh in 1963 and 1965; signs not written in grammatical order; suffixes, postpositions, verbal morphemes and the place determinative omitted; both editions describe their translations as provisional.](https://mysteriesdept.com/sources/SRC0DEDJQ13CJFX/zame-hymns-wikipedia-article)cited 3 times here**SRC** 0DEDJQ13CJFX
1. 24.[Cuneiform Digital Library Initiative n.d.. CDLI Terms of Use. cdli.earth.](https://mysteriesdept.com/sources/SRC0D1P8PQT7ZC2/cdli-terms-of-use)cited 7 times here**SRC** 0D1P8PQT7ZC2
1. 25.[Acta Sumerologica 4, pp. 1-6. Asks whether the UD.GAL.NUN sign substitutions of the Fara and Abu Salabikh literary texts hide an early form of the Emesal dialect; one of the competing explanations of that orthography. Used for the UD.GAL.NUN problem.](https://mysteriesdept.com/sources/SRC0DEDJ7JQW301/emesal-in-early-dynastic-sumerian-what-is-the-ud-gal-nun)cited 4 times here**SRC** 0DEDJ7JQW301
1. 26.[Used only as a map to the specialist literature on the lexical tradition: archaic word lists of the early third millennium with nouns only, the profession list Lú A copied almost unchanged for about a thousand years, the Fara god list, and the archaic list AD-GI4 (misnamed "Tribute") whose language has been called possibly non-Sumerian. Veldhuis's History of the Cuneiform Lexical Tradition (2014) is the standard account.](https://mysteriesdept.com/sources/SRC0DEDJPHPCC21/cuneiform-lexical-lists-wikipedia-article)cited 2 times here**SRC** 0DEDJPHPCC21
1. 27.[Zeitschrift für Assyriologie 61(2), pp. 193-207. Edition of the Early Dynastic IIIa copies of the Kesh Temple Hymn from Abu Salabikh (c. 2600 BC), about eight centuries older than the Old Babylonian copies and close to them except in spelling; Biggs notes the formidable obstacles these early literary texts put to comprehension. Used for the one Early Dynastic literary text that can be read through its later parallel.](https://mysteriesdept.com/sources/SRC0DEDJ4RSV9CP/an-archaic-sumerian-version-of-the-kesh-temple-hymn-from)[doi](https://doi.org/10.1515/zava.1971.61.2.193)cited 6 times here**SRC** 0DEDJ4RSV9CP
1. 28.[EMNLP 2021 (arXiv 2109.04513). Masked language models (multilingual BERT fine-tuned on Akkadian transliterations) proposing words for gaps in broken Akkadian texts, evaluated with held-out masking and with Assyriologists. Used for the gap-filling line of work after fetaya2020restoration.](https://mysteriesdept.com/sources/SRC0DEDGVD155EN/filling-the-gaps-in-ancient-akkadian-texts-a-masked-language)[doi](https://doi.org/10.18653/v1/2021.emnlp-main.384)cited 2 times here**SRC** 0DEDGVD155EN
1. 29.[COLING 2020, pp. 3454-3460, CC BY 4.0. First machine translation of Sumerian (Ur III administrative transliterations) into English from the CDLI Machine Translation and Automated Analysis of Cuneiform Languages (MTAAC) project; releases parallel corpora and compares supervised, phrase-based and transfer-learning systems. Used for the Sumerian side of machine translation.](https://mysteriesdept.com/sources/SRC0DEDH7RBVFFS/towards-the-first-machine-translation-system-for-sumerian)[doi](https://doi.org/10.18653/v1/2020.coling-main.308)cited 3 times here**SRC** 0DEDH7RBVFFS
1. 30.[GitHub repository behind gutherz2023akkadian. Python package (pip install akkadian) that transliterates Unicode cuneiform with HMM (89.5 per cent), MEMM (94 per cent) and BiLSTM (96.7 per cent) sign-reading accuracy on the RINAP training corpus; ships RINAP, RIAo, RIBo, SAAo and Suhu in JSON and TEI. README states CC BY-SA 3.0 while the sidebar shows MIT; licence to be confirmed from the licence file. Used as the open code and data behind the Akkadian machine-translation work.](https://mysteriesdept.com/sources/SRC0DEDGC4Z30M9/akkademia-translating-akkadian-signs-to-transcriptions-using)cited 3 times here**SRC** 0DEDGC4Z30M9
1. 31.[arXiv 2606.22608, submitted 21 June 2026, not yet peer reviewed. Deformable DETR sign detector at 173 and 106 sign classes, lines grouped and linked to transliterations by n-gram similarity; applied to 87,668 eBL tablet fragments giving about 2.9 million sign detections; authors note sensitivity to damage and layout and the absence of linguistic priors. Used for the current state of cuneiform OCR at scale.](https://mysteriesdept.com/sources/SRC0DEDGCC49XMN/automated-sign-detection-across-the-electronic-babylonian)cited 7 times here**SRC** 0DEDGCC49XMN
1. 32.[heiDATA deposit: frontally aligned 3D models with precomputed surface features and six-view raster images of tablets in the Hilprecht Collection, Jena, with metadata, transcriptions and transliterations for 707 tablets and a script to pull CDLI metadata; the parent Hilprecht Archive Online holds 1,977 3D scans. Open access, specific licence not stated on the record. Used as the reference 3D-scan benchmark for cuneiform.](https://mysteriesdept.com/sources/SRC0DEDH88H54WJ/heicubeda-hilprecht-heidelberg-cuneiform-benchmark-dataset)[doi](https://doi.org/10.11588/DATA/IE8CCN)cited 2 times here**SRC** 0DEDH88H54WJ
1. 33.[Cuneiform Digital Library Initiative n.d.. CDLI bulk data (daily dump of catalogue and transliterations). GitHub cdli-gh/data.](https://mysteriesdept.com/sources/SRC0D1P8PZM6DZE/cdli-bulk-data-daily-dump-of-catalogue-and-transliterations)cited 7 times here**SRC** 0D1P8PZM6DZE
1. 34.[Dataset card: 91,606 Sumerian tablets as Unicode glyph to transliteration pairs built from ePSD2/ORACC data, CC BY 4.0, splits 82,452 / 4,577 / 4,577, with period and genre splits; companion paper at the ML4AL workshop, ACL 2024. Used as the largest open Sumerian glyph-transliteration dataset.](https://mysteriesdept.com/sources/SRC0DEDH7WDK85B/sumtablets-a-sumerian-transliteration-dataset-hugging-face)cited 2 times here**SRC** 0DEDH7WDK85B
1. 35.[Hermalin, N. 2023. A Mutual Information-based Approach to Quantifying Logography in Japanese and Sumerian. Proceedings of the Workshop on Computation and Written Language (CAWL 2023).](https://mysteriesdept.com/sources/SRC0D8MR7Y219SS/a-mutual-information-based-approach-to-quantifying)[doi](https://doi.org/10.18653/v1/2023.cawl-1.12)cited 2 times here**SRC** 0D8MR7Y219SS
1. 36.[CDLI record for one Early Dynastic IIIa (c. 2600-2500 BC) witness of the Kesh Temple Hymn, excavation number AbS-T 0284, National Museum of Iraq IM 070193, with an approved ATF transliteration of three obverse columns and lacunae; composite Q000751, ETCSL 4.80.2. Used as the open record of an Early Dynastic literary tablet.](https://mysteriesdept.com/sources/SRC0DEDJNSNBQX9/cdli-literary-000751-ex-011-p010248-kesh-temple-hymn-witness)cited 4 times here**SRC** 0DEDJNSNBQX9

Last edited 31 Oct 2026