SumTablets
SumTablets turns Sumerian transliteration into a supervised
sequence-to-sequence task over Unicode glyphs, with open data and
reproducible baselines.
-
Corpus size
-
91,606 tablets /
6,970,407 glyphs
-
Source coverage
-
Readings to glyph 99.93%; glyph to
Unicode 99.96%
-
Evaluation
-
Character-level chrF, averaged by
tablet on held-out test data
-
Model delta
-
Weighted dictionary 61.22 →
neural baseline 97.54 chrF
The dataset pairs source glyph sequences with transliterations for
91,606 tablets (about
6.97M glyphs), preserving structural
context via special tokens for surfaces, line breaks, rulings, columns,
blank space, and breakage.
On a stratified 90/5/5 split by
historical period, a weighted dictionary sampler reaches
61.22 chrF, while an XLM-R-initialized
encoder-decoder reaches 97.54 chrF. The
objective is practical: speed philologist review, target uncertain
readings, and make downstream restoration/translation pipelines tractable.
The bottleneck was not model choice first. It was a stable, machine-usable
pairing of glyph strings with transliteration strings that preserves
tablet structure and Assyriological conventions.
Input glyphs
𒀀𒂗𒆤 𒈦𒆳𒆳𒊏
𒀊𒁀𒀀𒀀𒌷𒉈𒆤
...
Output transliteration
{d}en-lil2 lugal kur-kur-ra ab-ba dingir-dingir-re2-ne-ke4 ...
Step 01 / Ingest + Normalize
Standardize transliterations from ePSD2/Oracc pipelines
Source transliterations are cleaned and canonicalized into a consistent
format suitable for glyph-to-text learning, while retaining period and
genre metadata for each tablet ID.
Step 02 / Reading → Glyph Mapping
Recover source glyphs from transliterated readings
A two-stage mapping converts transliterated readings to glyph names and
then glyph names to Unicode cuneiform signs. Successful conversion
retained 6,724,498 readings (99.93%) and 6,638,081 Unicode glyph
mappings (99.96%).
Step 03 / Structural Fidelity
Encode tablet layout as aligned special tokens
Structure tokens are inserted in both glyph and transliteration streams
so models do not lose tablet-level formatting that affects reading
constraints.
[<SURFACE>] [
] [...] [<RULING>] [<COLUMN>] [<BLANK_SPACE>]
-
Periods
-
10 labeled periods for temporally
aware evaluation and transfer studies.
-
Genres
-
14 labeled genres exposing
style/domain shift across administrative, literary, legal, and ritual
texts.
-
Partitioning
-
90/5/5 train/val/test, stratified by
period to reduce period leakage and stabilize genre mix.
-
Lexical handling
-
Lexical texts excluded from validation/test and added to training only.
Two baselines establish the floor and ceiling: a weighted reading sampler
and a multilingual transformer encoder-decoder with task-specific
tokenization.
Baseline A / Non-neural
Weighted dictionary sampler
For each glyph, sample from known readings proportional to observed
frequency. The weighted mean number of readings per glyph is
6.75.
This baseline establishes how far frequency-only disambiguation can go
without context modeling.
[lookup sampling] [frequency prior] [no sequence context]
Baseline B / Neural
XLM-R initialized encoder-decoder
Encoder and decoder both initialize from a
279M-parameter XLM-R checkpoint,
then are adapted for glyph-to-transliteration generation.
Custom SentencePiece vocabularies:
632 glyph tokens and
1024 transliteration tokens, each
including 11 shared special tokens.
[XLM-R] [SentencePiece] [seq2seq] [beam search]
-
Stage 1
-
Encoder MLM pretraining: 50 epochs,
seq len 64, LR
5e-5, batch
2048, mask prob
0.10, warmup
200.
-
Stage 2
-
Joint model with frozen encoder: LR
1e-4,
2 epochs, batch
128, warmup
100.
-
Stage 3
-
Joint model with unfrozen encoder: LR
5e-5,
4 epochs, batch
128, warmup
100.
-
Inference + Compute
-
Beam size 5, AdamW optimization,
trained on single A100 SXM 80GB.
The neural baseline closes most of the transliteration gap, but
performance remains uneven across historically and stylistically distinct
genres.
Overall
-
Dictionary
- 61.22
-
Neural
- 97.54
Administrative
-
Dictionary
- 63.15
-
Neural
- 98.14
Royal inscription
-
Dictionary
- 54.58
-
Neural
- 95.15
Literary
-
Dictionary
- 37.73
-
Neural
- 90.67
Liturgy
-
Dictionary
- 55.92
-
Neural
- 77.68
Interpretation
Genre imbalance still matters
Administrative texts dominate training distribution and also yield the
highest scores. Lower-resource genres (especially liturgical/literary)
remain a meaningful error source, suggesting gains from balanced
curriculum sampling and uncertainty-aware decoding.
Operational Value
Useful now as expert-in-the-loop tooling
Even with residual errors, high chrF at corpus scale makes the model
practical for draft transliteration generation and triaging legacy
transliterations for specialist review rather than full manual
first-pass reading.
SumTablets is infrastructure: it turns transliteration from a one-off
manual act into a reproducible ML problem with public benchmarks and
reusable assets.
Research trajectory
From transliteration to restoration and translation
A strong glyph-to-transliteration model is a prerequisite for
damaged-tablet restoration, uncertainty estimation over readings, and
low-resource Akkadian/Sumerian translation pipelines.
Scholarly unlock
Scale philological workflows without flattening nuance
By preserving layout and metadata, the dataset supports models that can
respect period, genre, and document structure rather than collapsing
tablets into context-free token streams.
My contribution
End-to-end technical ownership
Led data design, preprocessing, Unicode mapping, split strategy,
tokenizer retraining, baseline implementation, training, evaluation, and
release packaging across ACL paper, repository, and Hugging Face assets.