📄 From Tablet to Text: AI-Assisted Reconstruction and Translation of Akkadian Passages in the Epic of Gilgamesh

Summary

The paper looks at how far neural translation models can help philologists restore and translate damaged Akkadian lines of the Epic of Gilgamesh, and where they stop being useful. The authors assembled a digital corpus of Standard Babylonian tablets with Old Babylonian parallels. It combines Unicode cuneiform and tablet images, normalised Latin transliterations, and several published English translations aligned line by line. They fine-tuned a moderately sized transformer NMT model in two configurations, one from signs and one from transliteration. They connected it to an interactive interface where experts can accept, reject or edit the model's ranked restorations. The evaluation used three tests: artificially masked lines with known readings, genuinely disputed lines, and related Akkadian texts outside the epic. On masked lines, the exact original appeared among the top five proposals about 78% of the time, and accuracy was highest on formulaic material. Rare words, broken names and syntactically unclear lines produced fluent but unsupported output. The authors therefore present AI as a supervised junior collaborator, not an authority. They also stress Iraqi leadership and explicit labelling of machine-assisted readings.

Key points

• Two model setups: C2E (Unicode cuneiform signs → English) and T2E (transliteration → English). Both use SentencePiece subword tokenisation and transfer learning from morphologically rich Semitic languages.
• Several English translations were included for each line where possible, so the model would not copy a single editor's style. Tablet and line identifiers were kept so every output can be traced to a physical object.
• Test 1: 100 undisputed lines were partly masked, at the level of signs, phrases or half-verses. The exact original was in the top-5 suggestions in roughly 78% of cases, with the best results on repeated epithets and narrative transitions.
• Test 2: 30 contested lines, mainly from Tablets II, V and XI. Experts compared each output with previously published restorations and judged whether it matched one of them, offered a variation on one, or departed from all of them.
• Test 3: companion texts such as Atrahasis measured whether the model had learned Gilgamesh-specific phrasing or more general Akkadian epic style. The main failure modes were rare lexemes, damaged proper names and uncertain syntax, plus confident hallucinations.

Interesting ideas

• I1. Hide parts of inputs whose correct answer is already settled and score recovery in the top-k; this gives a ground-truth ceiling before testing on cases where no answer is known.
• I2. When there is no single right answer, sort system output against the existing human alternatives as matching, a variant, or divergent, instead of forcing a right/wrong score.
• I3. Train and score against several independent references so the system does not simply imitate one source's style.
• I4. Run the system on related texts outside the training material to see whether it learned general patterns or memorised the style of its corpus.
• I5. Mark every machine-proposed item with its probability, the alternatives it beat, and a link back to the source evidence, and record each accept or reject decision with its reasoning.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Mask words or arguments in sentences that already have trusted gold parses and frames, and check whether Mova restores them. This would serve as a calibrated control before trusting its silver output on harder text.
• I2 (tree): Where treebanks or UD dialects disagree on an attachment, label each parser's choice as one of the accepted analyses, a variant, or an outlier. This keeps legitimate ambiguity from being counted as an error.
• I3 (back): Judge text generated from the meaning graph against multiple human paraphrases rather than one target sentence. This rewards preserving meaning over copying surface wording.
• I4 (answer): Give the readers tales and tests from collections outside the familiar fable set. A drop in accuracy would expose answers that depend on memorised phrasing instead of an understood text world.
• I5 (world): Store inferred entities, states and event orderings as labelled hypotheses tied to text spans, with ranked alternatives. This keeps every world fact explainable and lets later stages revise it.