💬 NLP

TabletCraft: Bridging a 4,000-Year Cultural Gap with Bidirectional Akkadian NMT and Cuneiform Rendering

TabletCraft is the first open-source, bidirectional system that bridges the 4,000-year cultural gap with ancient Mesopotamia by enabling users to translate between Akkadian and English and render new messages as cuneiform clay tablets, achieving competitive BLEU scores on both translation directions.

Original authors: Zhaohui Wang

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhaohui Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where you can read a letter from a king who lived 4,000 years ago, but you can't write one back. That's the strange reality of our oldest writing system, cuneiform. For centuries, this system has been locked behind a wall of time, accessible only to a tiny group of experts who can decipher the wedge-shaped marks pressed into clay tablets. Think of it like a library where everyone can walk in and read the books, but the doors are locked from the inside, and no one knows how to make a new key. This isn't just about translating words; it's about a one-way street in history. We can decode the past, but we can't speak to it. This paper sits at the intersection of computer science and ancient history, tackling the challenge of "Neural Machine Translation" (NMT)—where computers learn to translate languages by finding patterns in massive amounts of text—and applying it to a language that has been dead for millennia. The big question isn't just "Can we translate it?" but "Can we write in it?" If we can't, we remain passive tourists in history; if we can, we become active participants.

Enter TabletCraft, a new open-source tool that tries to break down that 4,000-year-old wall. The researchers built a system that doesn't just translate ancient Akkadian into modern English; it does the reverse, too. You can type a sentence like "The king rules the land" in English, and the computer will translate it into ancient Akkadian, turn those words into the correct cuneiform symbols, and finally, render them onto a digital image of a clay tablet. It's like having a magical time-traveling pen that lets you write a postcard to the past.

The team trained a powerful computer brain (a model based on something called ByT5) on 116,000 pairs of sentences, teaching it to go both ways: from ancient to modern and from modern to ancient. They also built a massive digital dictionary with over 14,000 mappings to ensure that when the computer picks a word, it finds the right ancient symbol to represent it. The results are impressive: on a test set of nearly 3,000 samples, the system achieved a translation score of 49.1 when going from Akkadian to English, and 48.5 when going from English to Akkadian. This is the first time anyone has published a quantitative score for translating into Akkadian at this scale.

However, the author is very careful not to call this a perfect solution. They explicitly state that the English-to-Akkadian direction produces "approximate" translations, not authentic ancient texts. Because ancient Akkadian had many ways to spell the same thing (like using different symbols for the same sound), the computer might pick a valid but different version than the one used in the test data. This means the scores are likely a "lower bound," or a conservative estimate, of how good the system actually is. The paper also notes that the system works best with royal inscriptions from a specific time period (Neo-Assyrian) and might struggle with the slangy, abbreviated language of ancient merchant letters.

The goal of TabletCraft isn't to replace historians or to let people fake ancient documents. Instead, it's designed to let students, museum visitors, and curious people "play" with the script. A student could write their name on a digital clay tablet; a museum-goer could compose a short dedication. The system labels every machine-generated output as "approximate" to ensure no one mistakes it for a real historical artifact. By opening the door to writing in cuneiform, TabletCraft shifts the experience from just reading history to actively engaging with it, even if that engagement is a bit rough around the edges. It's a first step toward making the world's oldest writing system a living, two-way conversation rather than a silent, one-way archive.

Technical Summary: TABLETCRAFT

Problem Statement
Despite the survival of approximately 500,000 cuneiform clay tablets, a significant cultural barrier persists: modern users cannot read or write in the world's oldest writing system. Existing Natural Language Processing (NLP) tools, such as recent Neural Machine Translation (NMT) systems, address this only partially by enabling one-way translation from Akkadian to English. This creates a "temporal asymmetry" where modern users remain passive consumers of ancient culture. Crucially, no existing system allows non-specialists to compose new content in cuneiform or translates English back into Akkadian and renders it visually, limiting engagement to passive reading rather than active participation.

Methodology
The author presents TABLETCRAFT, an open-source system designed to bridge this gap through bidirectional interaction. The system architecture consists of three primary components:

Bidirectional Translation Model:
- Base Architecture: A fine-tuned ByT5-base model (581M parameters).
- Training Data: The model was trained on 116,000 bidirectional samples derived from 58,126 parallel sentences. These sources include the Akkademia dataset (50K), a shared translation task from the Deep Past Initiative (1.5K), and sentence-aligned expansions (6K).
- Training Strategy: The model is trained bidirectionally (Akkadian→English and English→Akkadian) using task prefixes. Byte-level tokenization is employed to handle diacritics (e.g., š, t., s.) and logograms without vocabulary mismatch.
- Configuration: Training utilized 20 epochs, a learning rate of $10^{-4}$ , batch size 32, label smoothing of 0.2, and BF16 precision on two RTX 3090 GPUs.
Cuneiform Sign Converter:
- This component acts as a deterministic lookup table mapping transliteration strings to Unicode cuneiform characters.
- It contains 14,240 mappings compiled from ORACC, CDLI, and Akkademia, achieving 95.3% token coverage.
- Following Assyriological convention, determinatives are stripped during the conversion process.
Visual Tablet Renderer:
- This module generates SVG/PNG outputs styled after Neo-Assyrian tablets, featuring a clay background, ruling lines, and the Noto Sans Cuneiform font.
- It operates without GPU acceleration, rendering a tablet in under 10 milliseconds.

The system is packaged as a pip-installable toolkit (cuneiscribe) with a CLI, Python API, and a Gradio web demo.

Key Contributions

First Bidirectional System: TABLETCRAFT is the first open-source system to support both Akkadian→English translation and English→Akkadian composition, coupled with visual cuneiform rendering.
Cultural Agency: By enabling English-to-cuneiform composition, the system reframes users from passive readers to active participants, allowing them to generate personal messages (e.g., names, dedications) in the ancient script.
Dual-User Design: The system addresses the tension between scholarly fidelity and public accessibility. It exposes intermediate layers (transliteration, sign mapping) and explicitly labels all machine-generated Akkadian as "approximate," serving both researchers (who need draft glosses for verification) and the general public (who seek personalized engagement).
Open Accessibility: The toolkit lowers barriers for researchers and educators in modern nations encompassing ancient Mesopotamia (Iraq, Syria, Turkey, Iran) by providing a multilingual-extensible, open-source resource.

Results
The system was evaluated on the held-out Akkademia validation split (2,812 samples, primarily Neo-Assyrian royal inscriptions):

Akkadian→English: Achieved 49.1 BLEU and 63.1 chrF++, outperforming the prior Akkademia CNN baseline (37.5 BLEU).
English→Akkadian: Achieved 48.5 BLEU and 55.6 chrF++. This represents the first published quantitative result for this direction.
- Note on Metrics: The author notes that BLEU scores for English→Akkadian are conservative lower bounds. Because Akkadian admits multiple philologically valid spellings (e.g., logographic vs. syllabic) for the same surface form, a fluent paraphrase using different valid signs may score lower despite being correct.
Coverage: The cuneiform converter achieved 95.3% token coverage on sampled transliterations.

Significance and Claims
The paper claims that TABLETCRAFT enables a "participatory mode of interaction" with cuneiform, moving beyond the passive consumption of ancient texts. It emphasizes that the system is designed as a "draft-and-revise assistant" rather than an authoritative translator. The author explicitly states that outputs should not be cited as primary historical evidence, as the English→Akkadian direction produces approximate modern transliterations rather than authentic ancient text.

The work highlights that "Akkadian" cannot be treated as a monolithic language; performance varies significantly by dialect and genre (e.g., Neo-Assyrian royal inscriptions vs. Old Assyrian commercial letters). Consequently, the system includes confidence scoring and genre tagging to manage user expectations. The paper concludes that while the system successfully bridges a 4,000-year cultural gap for composition, future work is required to support Sumerian, cuneiform OCR, and non-English language pivots.

Technical Summary: TABLETCRAFT

More like this