◆ NLP · Ancient Language

Reading Sumerian

Type transliterated Sumerian and get a word-by-word English reading. The dictionary behind it was not hand-written, it was learned by aligning 32,578 Sumerian and English phrases from real cuneiform tablets. This is the first, honest layer of a translator, a neural model is the next.

← case studies
Read this first

This is a first-pass gloss, not a scholarly translation. It reads transliterated Sumerian (Latin letters in the ATF style scholars use, not cuneiform pictures) using the corpus conventions: hyphens between signs (lugal-e), determinatives in braces ({d}nanna, lagasz{ki}), and sz for the "sh" sound. It has no grammar and no word order, it labels each word with its most likely meaning. Unknown words show in red. Sumerian is ambiguous (one reading can mean several things, and grammar words often fall back to a weak default like "the"), so treat this as a helper for a human, exactly how the researchers who built the data intend it.

Try it

Enter transliterated Sumerian. Numbers like 1(disz) or plain digits become [number]. Diacritics (š, ) are accepted and folded to ATF.
Or try a real line from a tablet:
English reading
Word by word
Phrase pairs learned from
--
Sumerian words known
--
Word coverage (held-out)
--
Meaning overlap vs scholars
--

How it learns a dictionary from nothing

Nobody typed in that udu means "sheep." The model saw tens of thousands of short phrase pairs and used IBM Model 1, a classic word-alignment algorithm: it starts by assuming every English word could translate every Sumerian word, then repeatedly rewards pairs that keep co-occurring. After ten passes, udu has locked onto "sheep," ki onto "from," lugal onto "king," kiszib3 onto "seal." It is the same idea as counting which words travel together, made precise.

The honest failure it exposes: Sumerian is full of ambiguity. The same reading can mean several things, and grammatical particles carry little meaning on their own, so the model falls back to a weak "the." Seeing that is the point, it is exactly where the next layer, a neural sequence-to-sequence model that reads a whole line in context, earns its keep.

That neural model is the next build, and it extends the same NLP pipeline this project already uses: text to integer sequences, padding to a uniform length, an embedding layer, then an LSTM encoder-decoder.

Data: MTAAC / CDLI Sumerian-English parallel corpus (all compositions), via the cdli-gh project. Lexicon learned with IBM Model 1 alignment. Built by Carlos Abel Vivanco / AbleV Labs · ablevlabs.com ↗