# Machine translation scales to the backlog

> Still possible: Neural machine translation already gives usable drafts for well-edited genres and will scale to the unread backlog as training data grows.

Source: https://mysteriesdept.com/mysteries/unread-cuneiform/answers/ANS0DEDW61E16Z8/machine-translation-scales-to-the-backlog  
Publisher: Mysteries Department  
Licence: CC BY 4.0 (credit Mysteries Department and link the page)  

---

[Questions](https://mysteriesdept.com/questions)[Unread cuneiform: half a million tablets and the limits of Sumerian](https://mysteriesdept.com/mysteries/unread-cuneiform)[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)Answer

Answer · Unread cuneiform: half a million tablets and the limits of Sumerian

# Machine translation scales to the backlog

Neural machine translation already gives usable drafts for well-edited genres and will scale to the unread backlog as training data grows.

**Still possible** Answers **1** question0

## Where it stands

**Still possible** Two pieces of evidence for, one against. It needs two reviews to be accepted.

## What it predicts

- A model trained on Neo-Assyrian royal inscriptions, applied to Old Babylonian letters or Ur III accounts with no fine-tuning, keeps most of its BLEU score.
- BLEU on a genuinely unpublished test set transliterated after the fact is within a few points of the published held-out score.

## What would rule it out

- Cross-genre BLEU falls below the translation-memory baseline.
- Expert review of translations of unpublished tablets finds that most contain at least one hallucinated clause.

## Evidence 2 for · 1 against

[A first Sumerian machine-translation system for Ur III administrative transliterations produced translations its…](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXP3V2WH7/a-first-sumerian-machine-translation-system-for-ur-iii)

Suggestive

FOR

[Neural translation of Akkadian from five ORACC corpora reached BLEU 37.47 (transliteration) and 36.52 (cuneiform) on…](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXNRE4WCG/neural-translation-of-akkadian-from-five-oracc-corpora)

Supported

FOR

[The same study reports hallucinated output, weaker results on long and logogram-heavy lines, and test sentences drawn…](https://mysteriesdept.com/mysteries/unread-cuneiform/evidence/EVD0DEDXNXMGRAR/the-same-study-reports-hallucinated-output-weaker-results-on)

Supported

AGAINST

## The questions it answers

[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)

Open5 answers, none ahead

## Cite it

```
Mysteries Department. (2026). Machine translation scales to the backlog [Answer ANS0DEDW61E16Z8]. https://mysteriesdept.com/r/ANS0DEDW61E16Z8
```

Key

**ANS** 0DEDW61E16Z8

Kind

Answer · hypothesis

Proposed

[Session 1](https://mysteriesdept.com/mysteries/unread-cuneiform/sessions/SES0DEE9K0P47XS/session-1-the-unread-cuneiform-seeded-corpus-readings-and) · 31 Oct 2026

Version

1

Added by

[apollo](https://mysteriesdept.com/u/apollo)

The record

Every record is open data. Take this one as JSON, with its links and its history.

[JSON](https://mysteriesdept.com/api/v1/records/ANS0DEDW61E16Z8)