SumTablets: A Transliteration Dataset of Sumerian Tablets

Cole Simmons, Richard Diehl Martinez, Dan Jurafsky

TL;DR

Researchers have released SumTablets, the first large-scale, standardized dataset pairing over 91,000 Sumerian cuneiform tablets with their transliterations. By fine-tuning a multilingual transformer model on this data, they achieved a staggering 97.54 chrF score, demonstrating that AI can now assist experts in rapidly deciphering the world’s earliest written language.

What Problem Does This Paper Address?

Sumerian is a language isolate written in cuneiform, a script consisting of wedge-shaped impressions on clay. For over 150 years, Assyriologists have practiced transliteration: converting these physical glyphs into a Latin-script representation that reflects their phonetic or semantic meaning.

The challenge lies in polyvalency. A single cuneiform sign can have many different "readings" depending on the context. For example, the sign KA can be read as ka ("mouth"), dug₄ ("to speak"), kiri₃ ("nose"), or zuh ("to steal"). Historically, digital archives like Oracc have published transliterations, but they rarely included the original digital glyphs (Unicode) alongside them. This lack of paired data made it impossible to train modern Natural Language Processing (NLP) models to automate or assist in the transliteration process.

Methods

The authors built SumTablets by "reverse-mapping" existing digital transliterations back to their original Unicode glyphs.

  1. Data Cleaning: They parsed JSON data from the Electronic Pennsylvania Sumerian Dictionary (ePSD2), removing modern scholarly annotations (like brackets for broken text) to focus on what was actually on the tablet.
  2. Mapping: Using sign lists, they converted each transliterated word back into its constituent Unicode cuneiform characters.
  3. Structural Tokens: They introduced special tokens to preserve the physical layout of the tablets, such as <SURFACE> (obverse/reverse), \n (line breaks), and <RULING> (horizontal lines drawn by scribes).

To test the dataset, they compared two approaches:

  • Dictionary Baseline: A simple probabilistic model that samples a reading for a glyph based on how often that reading appears in a dictionary.
  • Neural Baseline: An encoder-decoder model based on XLM-R, a large multilingual transformer. They fine-tuned the model using a multi-stage process, including Masked Language Modeling (MLM) on the cuneiform glyphs themselves.

Datasets

The primary contribution is the SumTablets dataset:

  • Size: 91,606 tablets containing nearly 7 million glyphs.
  • Content: Paired Unicode glyphs and transliterations, plus metadata for time period and genre.
  • Availability: Released via Hugging Face (CC BY 4.0).
  • Sources: Derived from Oracc, ePSD2, and the Oracc Sign List (OSL).

Key Insights and Ideas

  • Context is King: The dictionary-only approach failed (61.22 chrF) because it couldn't handle polyvalency. The neural model’s success (97.54 chrF) proves that transformers are excellent at using surrounding signs to disambiguate which "reading" of a glyph is correct.
  • Genre Matters: The model performed best on administrative texts. These are often formulaic (receipts, ledgers), making them more predictable than literary or liturgical texts, which have more varied vocabulary and style.
  • The "Nasal G" Problem: The authors noted that inconsistent scholarly conventions (e.g., writing "saŋ" vs. "sag") confuse the model. This highlights a need for better standardization in digital humanities data.
  • Transfer Learning for Isolates: Even though Sumerian is a language isolate (related to no known living language), leveraging a model pre-trained on 100+ modern languages (XLM-R) provided a powerful starting point for the task.

Key Results

The neural model significantly outperformed the dictionary baseline across all historical periods and genres:

Model Overall chrF Score
Dictionary Baseline 61.22
Neural Baseline (XLM-R) 97.54

In specific categories like the Ur III period (which makes up the bulk of the data), the model reached an accuracy of 98.46. Even in less-represented periods like the Neo-Babylonian, it maintained a high score of 97.81.

Limitations and Future Work

  • Data Imbalance: The dataset is heavily skewed toward the Ur III period and administrative genres. Performance on rare literary texts is lower and requires more targeted data.
  • Missing Glyphs: Some complex compound glyphs are not yet represented in the Unicode standard, forcing the model to use <UNK> (unknown) tokens.
  • Phonemes in Names: The model still struggles with proper names, where glyphs are often used purely for their sound rather than their meaning.
  • Future Goal: The authors envision this as a first step toward a full "Sumerian translation pipeline," moving from raw tablet images to Unicode glyphs, then to transliteration, and finally to English translation.

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