SumTablets: A Transliteration Dataset of Sumerian Tablets

Cole Simmons, Richard Diehl Martinez, Dan Jurafsky

TL;DR

SumTablets is the first large-scale, open-access dataset that pairs nearly 92,000 Sumerian cuneiform tablets with their digital transliterations. By fine-tuning a multilingual transformer model (XLM-R) on this data, researchers achieved a staggering 97.55 character-level F-score, proving that AI can effectively assist scholars in deciphering the world's oldest written language.

What Problem Does This Paper Address?

Sumerian, recorded in the cuneiform script on clay tablets, represents the transition from prehistory to history. For over 150 years, Assyriologists have performed "transliteration"—the process of interpreting these wedge-shaped glyphs into the Latin script.

The challenge is that cuneiform is polyvalent: a single glyph can have dozens of different readings (phonetic or semantic) depending entirely on its context. For example, one specific sign can mean "mouth," "to speak," "nose," or "to steal." While digital archives like Oracc have published thousands of transliterations, they rarely include the original digital glyphs (Unicode) in a parallel format. This lack of "glyph-to-transliteration" pairs has been a major bottleneck, preventing modern Natural Language Processing (NLP) from helping scholars automate or verify this painstaking work.

Methods

The authors created SumTablets by standardizing data from several major digital Assyriology projects.

  1. Data Alignment: They mapped transliterated "readings" back to their source Unicode glyphs using specialized dictionaries (ePSD2 and the Oracc Sign List).
  2. Structural Integrity: Unlike standard text datasets, they used special tokens to preserve the physical structure of the tablets, such as <SURFACE> (to indicate the front or back of a tablet), \n (line breaks), and ... (to mark broken or missing segments).
  3. The Neural Baseline: They leveraged XLM-R, a massive transformer model pre-trained on over 100 modern languages. They adapted it to Sumerian through a three-stage process:
  4. Masked Language Modeling (MLM): Teaching the model to "fill in the blanks" of cuneiform sequences to learn internal representations of the glyphs.
  5. Frozen Encoder Training: Training a decoder to predict transliterations while keeping the glyph-knowledge fixed.
  6. Full Fine-Tuning: Training the entire encoder-decoder stack to optimize the conversion.

Datasets

  • SumTablets (New): The primary contribution. It contains 91,606 tablets totaling nearly 7 million glyphs. It is stratified by historical period (e.g., Ur III, Old Babylonian) and genre (e.g., Administrative, Legal, Literary).
  • Source Projects: The data was derived and cleaned from the Open Richly Annotated Cuneiform Corpus (Oracc), the Electronic Pennsylvania Sumerian Dictionary (ePSD2), and the Cuneiform Digital Library Initiative (CDLI).
  • Availability: The dataset is hosted on Hugging Face under a CC BY 4.0 license.

Key Insights and Ideas

  • Context is King: A single glyph has an average of 22.17 possible readings. A simple frequency-based dictionary lookup only gets about 61% accuracy. The transformer’s ability to look at the entire sequence allows it to disambiguate these readings with near-perfect precision.
  • Transfer Learning Works for Isolates: Sumerian is a "language isolate" (it has no known related languages). Despite this, a model pre-trained on modern languages like English, Basque, and Japanese still provided a powerful foundation for learning Sumerian syntax and morphology.
  • Administrative Dominance: The vast majority of surviving Sumerian texts are administrative (receipts, ledgers, etc.). The model performs best on these "formulaic" texts but struggles more with rare genres like Liturgy, where data is scarce.
  • Standardization vs. Reality: The authors note that Unicode "flattens" the script. In reality, cuneiform signs changed drastically over 3,000 years, but the dataset treats them as uniform digital characters to make them machine-readable.

Key Results

The researchers compared their neural model against a "Dictionary Baseline" (which simply picks the most likely reading for a sign based on frequency).

Model Overall chrF (Character F-score)
Dictionary Baseline 61.22
Neural Model (XLM-R) 97.54

The neural model's performance was remarkably consistent across most time periods, though it peaked at 98.46 for the Ur III period (the most data-rich era) and dipped for rarer genres like Liturgy (77.68).

Limitations and Future Work

  • Data Imbalance: Because the dataset is dominated by administrative documents from the Ur III period, the model may be biased toward that specific "dialect" and vocabulary.
  • The "Unknown" Problem: Some complex compound glyphs are not yet represented in the Unicode standard. These are currently mapped to <UNK> tokens, losing some information.
  • Orthographic Variation: The model doesn't yet account for the visual evolution of signs over millennia; it sees the "digital" version of the sign, not the actual wedge impressions.
  • Future Directions: The authors hope this dataset will lead to "expert-in-the-loop" tools where AI generates a draft transliteration that an Assyriologist can quickly verify, rather than starting from scratch. They also see this as a stepping stone toward a full Sumerian-to-English translation pipeline.

Summary updated

Abstract

Sumerian transliteration is a conventional system for representing a scholar's interpretation of a tablet in the Latin script. Thanks to visionary digital Assyriology projects such as ETCSL, CDLI, and Oracc, a large number of Sumerian transliterations have been published online, and these data are well-structured for a variety of search and analysis tasks. However, the absence of a comprehensive, accessible dataset pairing transliterations with a digital representation of the tablet's cuneiform glyphs has prevented the application of modern Natural Language Processing (NLP) methods to the task of Sumerian transliteration. To address this gap, we present SumTablets, a dataset pairing Unicode representations of 91,606 Sumerian cuneiform tablets (totaling 6,970,407 glyphs) with the associated transliterations published by Oracc. We construct SumTablets by first preprocessing and standardizing the Oracc transliterations before mapping each reading back to the Unicode representation of the source glyph. Further, we retain parallel structural information (e.g., surfaces, newlines, broken segments) through the use of special tokens. We release SumTablets as a Hugging Face Dataset (CC BY 4.0) and open source data preparation code via GitHub. Additionally, we leverage SumTablets to implement and evaluate two transliteration baselines: (1) weighted sampling from a glyph's possible readings, and (2) fine-tuning an autoregressive language model. Our fine-tuned language model achieves an average transliteration character-level F-score (chrF) of 97.55, demonstrating the immediate potential of transformer-based transliteration models in allowing experts to rapidly verify generated transliterations rather than manually transliterating tablets one-by-one.
BibTeX
@misc{simmons2026sumtablets,
  author = {Cole Simmons and Richard Diehl Martinez and Dan Jurafsky},
  title = {{SumTablets: A Transliteration Dataset of Sumerian Tablets}},
  year = {2026},
  eprint = {2602.22200},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.22200},
}
Figure