TabletCraft

The Problem: Cuneiform Is Readable, But Not Writable

Roughly half a million cuneiform clay tablets survive in museums worldwide, and the writing system is 4,000 years old. Existing NLP work enables one-way, scholar-oriented translation from Akkadian into English. That leaves non-specialists as passive consumers: you can have an ancient tablet explained to you, but you cannot write anything in cuneiform yourself. The reverse direction had no published quantitative baseline at all.

问题：现存楔形文字泥板约五十万块，但已有的 NLP 工作只做单向的「阿卡德语→英语」学术翻译。普通人只能被动接受释读结果，无法用楔形文字书写；反方向此前没有任何公开的定量基准。

Situation & Task: Make the Writing System Bidirectional

The task was to build a complete pipeline in both directions — read a tablet (Akkadian→English) and compose one (English→Akkadian→cuneiform signs→a rendered clay tablet image) — and to package it so that a non-specialist can install and run it without a research environment. A system that hands confident-looking cuneiform to someone who cannot read it carries a specific risk: the user has no way to detect that it is wrong, so silent failure is worse here than in ordinary MT.

任务：构建双向完整流水线——既能读懂泥板（阿卡德语→英语），也能创作泥板（英语→阿卡德语→楔形符号→渲染图像），并打包成非专业用户无需科研环境即可安装运行的工具。由于用户无法自行判断输出对错，「静默出错」在这里比普通机器翻译危险得多。

Innovation: Confidence Gating That Degrades Instead of Guessing

The pipeline never confidently renders wrong cuneiform. An input classifier runs before translation and an output validator runs after it. Inputs are typed (short / modern / anomalous) with a confidence score: modern concepts with no Akkadian equivalent raise an explicit warning, and anomalous input — including injection-shaped strings — is rejected outright rather than translated. When output quality is uncertain the system degrades gracefully to transliteration only, with the caveat attached, instead of emitting signs it cannot stand behind.

创新点：置信度门控。翻译前有输入分类器、翻译后有输出校验器，输入被分为 short / modern / anomalous 并给出置信度：无对应阿卡德语的现代概念会显式告警，异常输入（含注入型字符串）直接拒绝。输出不可靠时降级为「仅转写」并附带说明，而不是给出无法负责的符号。

Approach: Four Decoupled Layers

A ByT5-base bidirectional translator trained on 116K bidirectional samples handles the language pair; a knowledge layer holds 14,240 transliteration-to-Unicode sign mappings at 95.3% coverage; a renderer produces SVG/PNG clay tablets with Mesopotamian styling in under 10 ms; and an interface layer exposes a CLI, a Python API and a Gradio web demo. The layers are decoupled deliberately, so the model can be swapped, the sign tables updated, or a dialect added without breaking the interfaces.

方法：四层解耦架构。ByT5-base 双向翻译模型（116K 双向样本训练）、知识层（14,240 条转写→Unicode 映射，覆盖率 95.3%）、渲染层（SVG/PNG 泥板，10 ms 内完成）、接口层（CLI、Python API、Gradio 演示）。解耦是刻意的：换模型、更新符号表或增加方言都不会破坏接口。

Results and Significance

On the held-out Akkademia validation split (2,812 samples) the system reaches 49.1 BLEU for Akkadian→English and 48.5 BLEU for English→Akkadian — the first published quantitative result in the reverse direction, which is the contribution the paper rests on. It is released under Apache 2.0 on PyPI, so the cultural-access claim is testable by anyone rather than asserted. The paper is deliberate about limits: English→Akkadian yields approximate modern transliterations rather than authentic ancient text, training data is Neo-Assyrian and Old Babylonian so other dialects degrade, and the honest finding is that domain-specific data matters more than model size.

结果与意义：在 Akkademia 留出验证集（2,812 条）上，阿→英 49.1 BLEU、英→阿 48.5 BLEU，后者是该方向首个公开定量结果，也是论文的核心贡献。项目以 Apache 2.0 发布到 PyPI，使「降低文化门槛」这一主张可被他人验证而非仅是宣称。论文对局限也直言：英→阿产出的是近似的现代转写而非真实古文，训练数据集中于新亚述与古巴比伦方言，且真正的结论是「领域数据比模型规模更重要」。

• First published quantitative result for English→Akkadian: 48.5 BLEU, alongside 49.1 BLEU Akkadian→English on the Akkademia held-out split (2,812 samples)
• Confidence gating: input classifier + output validator; degrades to transliteration-only with a warning rather than rendering cuneiform it cannot stand behind
• 14,240 transliteration→Unicode sign mappings at 95.3% coverage
• Clay-tablet renderer (SVG/PNG) with Mesopotamian styling in under 10 ms
• ByT5-base bidirectional translator trained on 116K bidirectional samples
• Shipped as pip install cuneiscribe — CLI, Python API and Gradio web demo, Apache 2.0
• Anomalous input, including injection-shaped strings, is rejected rather than translated
• Accepted at C3NLP 2026 @ ACL (San Diego, July 2026), archival short paper