The Sumerian Verbal Chain is a Nightmare
I've been studying Sumerian for six months. Every week I think I understand the verbal system. Every week the next text proves me wrong. This post is my attempt to write down what I think I know, mostly so I can look back later and see how wrong I was.
The basic problem
Sumerian is agglutinative. That means grammatical information gets packed into a word by stacking affixes — prefixes, infixes, and suffixes — onto a verbal root. English does a little of this (un-help-ful-ness), but Sumerian does it with the enthusiasm of a language that never learned restraint.
A single Sumerian finite verb can encode:
- Modal prefix (is this a statement, a command, a wish?)
- Conjugation prefix (marks tense/aspect — roughly)
- Ventive/middle marker (direction of action toward/away from speaker)
- Coordinator (links clauses)
- Indirect object (dative)
- Comitative ("with")
- Ablative ("from")
- Causative
- Direct object (accusative)
- Verbal root
- Tense/aspect suffix
- Subject agreement (ergative or absolutive depending on transitivity)
That's twelve potential slots in a single word. Not all are filled at once, but enough of them are that a common Sumerian verb looks like someone dropped a bag of morphemes on the floor and glued them together in the order they landed.
A concrete example
Take this verb from a Gudea inscription:
mu-na-ni-in-du₃
Breaking it down:
| Slot | Morpheme | Function |
|---|---|---|
| Conjugation prefix | mu- | finite, ventive base |
| Dative (3sg) | -na- | "for him/her" |
| Locative-terminative | -ni- | "into it" |
| Subject (3sg ergative) | -in- | "he/she (did)" |
| Root | du₃ | "to build" |
So: "He built it for him (into it)." Five morphemes. One word. And I've simplified — the interaction between mu- and the ventive is actually a matter of ongoing scholarly debate that has produced at least three incompatible analyses since 1990.
Why it's hard for NLP
If you're coming from an NLP background, you might think: agglutinative language, use a morphological analyzer, done. Turkish is agglutinative and we have decent tools for that.
The problems specific to Sumerian:
No native speakers. Sumerian has been dead for about 4,000 years. There is no intuition to fall back on. When two Sumerologists disagree about a verb form, there is no way to ask a speaker who's right. The arguments are purely philological.
The writing system is ambiguous. Cuneiform signs can have multiple readings. The sign 𒆕 can be read as DU₃ ("to build"), TÙ ("to be born"), or several other values depending on context. The transliteration du₃ already represents an interpretive choice — the raw sign is underdetermined.
Small corpus, high variation. We have maybe 100,000 Sumerian texts, spanning about 1,500 years. The language changed significantly over that period. A verb form in an Ur III administrative text (2100 BCE) doesn't necessarily work the same way as the same form in an Old Babylonian literary text (1800 BCE). Any model needs to account for diachronic variation, which means splitting an already small dataset into even smaller period-specific subsets.
Slot interactions. The morpheme slots aren't independent. The conjugation prefix affects which object markers can appear. Transitivity determines the subject agreement pattern. The ventive interacts with everything. You can't just build a slot-filler; you need something that understands the constraints between slots.
What I'm trying
My current approach is a sequence-to-sequence model that takes a transliterated verb form as input and produces a morphological parse as output. The training data comes from the Electronic Text Corpus of Sumerian Literature (ETCSL), which has about 400 texts with morphological annotation.
Early results are... mixed. The model handles regular forms well — it can parse a clean transitive verb with standard agreement. But it falls apart on:
- Irregular prefix combinations (the mu-/ba-/im- alternations)
- Nominalized verbs (which have different morphology but look similar)
- Broken or ambiguous forms (which make up maybe 15% of the corpus)
The error rate on irregular forms is about 40%, which is terrible. For comparison, a first-year Sumerian student with a grammar book could probably do better on those same forms. The model has learned the regular patterns but not the exceptions, and in Sumerian the exceptions are where all the interesting linguistics happens.
The philosophical bit
There's a deeper question here that I keep coming back to: what does it mean to "understand" a dead language? When I parse a Sumerian verb, I'm applying a grammatical framework that was developed by European scholars in the 20th century and retrofitted onto texts written 4,000 years earlier. The Sumerians didn't think in terms of "ergative-absolutive alignment" or "conjugation prefixes." Those are our categories, imposed on their language.
A model trained on our annotations learns our framework, not the language itself. If our framework is wrong — and there are serious scholars who think parts of it are — then the model will learn the wrong thing with great confidence.
This doesn't mean the effort is pointless. A morphological parser, even an imperfect one, saves researchers enormous amounts of time. Manually parsing every verb in a new text takes hours. An automated first pass that's right 80% of the time and flags uncertain forms for human review would be genuinely useful.
But I try to hold the uncertainty honestly. We are not decoding Sumerian the way we decode a substitution cipher, with a single correct answer waiting to be found. We are building interpretive models of a language that exists only in ambiguous inscriptions on clay. Every parse is a hypothesis. The best we can do is make the hypotheses as good as possible and be transparent about where they're shaky.
That's what I'm working toward. More on this as the model improves — or fails instructively.