1 / 27
Restoring ancient texts修复古代文本
Google DeepMind · Nature 603 (2022)Google DeepMind · Nature 603 (2022)
谷歌 DeepMind · 《自然》603 卷（2022）谷歌 DeepMind · 《自然》603 卷（2022）
IthacaIthaca
How a deep neural network learned to restore, place, and date ancient Greek inscriptions — explained from the ground up, for everyone.一个深度神经网络如何学会修复、定位与定年古希腊铭文 —— 从零讲起，面向所有人。
62%62%
restoration, alonerestoration, alone
独立修复独立修复
25→72%25→72%
human + Ithacahuman + Ithaca
人类 + Ithaca人类 + Ithaca
71%71%
place, 84 regionsplace, 84 regions
定位（84 区）定位（84 区）
<30 yr<30 yr
dating errordating error
定年误差定年误差
▶ The full journey (merged deck)▶ The full journey (merged deck)
▶ 完整旅程（合并演示）▶ 完整旅程（合并演示）
2 / 27
A · The problemA · 问题
A · the problemWriting that survived — but brokenA · 问题幸存的文字 —— 却已残破
An inscription is text the ancients cut into stone. Over centuries they crack, erode, and lose their date — posing three tasks at once.一条铭文是古人刻入石头的文字。历经数百年，它们开裂、风化、年代失考 —— 同时抛出三项任务。
What did the lost letters say? — restoration缺失字母原本写了什么？ —— 修复
Where & when was it written? — attribution (no radiocarbon for stone)何地、何时书写？ —— 归属（石头无放射性碳）
Why it mattersThe old toolkit is slow: literal string search misses anything phrased differently and can’t give a probability over all readings.为何重要旧工具很慢：字面字符串检索会漏掉任何措辞不同者，也无法给出各读法的概率。
3 / 27
A · A worked exampleA · 一个实例
A · see it for yourselfA broken signatureA · 亲眼一看残缺的署名
I.PHI #315181 — the sculptor’s verb εποιη[σε] (“made it”) is damaged. A human restoring it alone is right ~25% of the time. Run Ithaca and watch it propose ranked guesses — filling the most confident position first. Assael 2022I.PHI #315181 —— 雕刻者的动词 εποιη[σε]（“制作”）残损。人类独立修复约 25% 正确。运行 Ithaca，看它给出排名猜测 —— 先填最有把握的位置。Assael 2022
▶ interactive: restore — open the live deck to use it交互演示：restore —— 打开实时演示以使用
Why it mattersIthaca never returns one verdict — it returns the top 20 ranked hypotheses so the scholar makes the final call.为何重要Ithaca 从不给出唯一定论 —— 它返回排名前 20 的假设，让学者做最终判断。
4 / 27
B · Building I.PHIB · 构建 I.PHI
B · the dataFrom a messy library to a clean textbookB · 数据从杂乱图书馆到整洁课本
The team turned the PHI corpus of 178,551 transcribed inscriptions into I.PHI: 78,608 cleaned, labelled examples. Assael 2022团队把含 178,551 条转写的 PHI 语料，转化为 I.PHI：78,608 条经清洗、带标注的样本。Assael 2022
▶ interactive: funnel — open the live deck to use it交互演示：funnel —— 打开实时演示以使用
5 / 27
B · 84 regionsB · 84 个区域
B · labelsEvery text tagged: where & whenB · 标注每条文本都被标注：何地、何时
Each inscription is tagged to one of 84 regions. The data is steeply uneven — Attica dominates — which is why label smoothing matters. Assael 2022每条铭文被标注到 84 个区域之一。数据极不均衡 —— 雅典占多数 —— 这正是标签平滑之故。Assael 2022
▶ interactive: treemap — open the live deck to use it交互演示：treemap —— 打开实时演示以使用
6 / 27
C · Text → numbersC · 文本 → 数字
C · representationHow do you feed a poem to a machine?C · 表示怎样把一首诗喂给机器？
A network only eats numbers. Every character becomes an embedding — and for each position three streams combine: character, word, and position. Assael 2022网络只“吃”数字。每个字符变成一个嵌入 —— 每个位置三股流合并：字符、词、位置。Assael 2022
Why it mattersOpen the related → embedding chip for the live three-streams demo (in the Concept Lab). A missing letter is posed as a “–” the model must fill.为何重要点击相关 → 嵌入标签看实时三流演示（原理实验室）。缺失字母以“–”提出，由模型填补。
7 / 27
Foundations · Transformer基础 · Transformer
FoundationsThe engine: a Transformer基础引擎：一个 Transformer
Ithaca’s core is a transformer (2017) — the same family behind modern language AI. Step through what it does. Vaswani 2017Ithaca 的核心是一个 transformer（2017）—— 与现代语言 AI 同源。逐步看它做什么。Vaswani 2017
▶ interactive: transformer — open the live deck to use it交互演示：transformer —— 打开实时演示以使用
8 / 27
Foundations · Attention基础 · 注意力
FoundationsAttention: who looks at whom基础注意力：谁在看谁
Every character scores how relevant every other is, then mixes in the important ones. Hover any letter; drag the slider through the 8 layers. Vaswani 2017每个字符为其他每个的相关度打分，再融合重要者。悬停任一字母；拖动滑块走过 8 层。Vaswani 2017
▶ interactive: attention — open the live deck to use it交互演示：attention —— 打开实时演示以使用
9 / 27
Foundations · BigBird基础 · BigBird
Foundations · prior technologyBigBird: pay attention, cheaply基础 · 既有技术BigBird：便宜地“关注”
Ithaca’s torso is 8 blocks of BigBird-style sparse attention — global + window + random — cutting cost from n² to ~linear. Toggle them. Zaheer 2020Ithaca 主干是 8 个 BigBird 式稀疏注意力块 —— 全局 + 窗口 + 随机 —— 把代价从 n² 降到近似线性。切换它们。Zaheer 2020
▶ interactive: bigbird — open the live deck to use it交互演示：bigbird —— 打开实时演示以使用
10 / 27
D · Torso + headsD · 主干 + 头
D · architectureOne trunk, three specialist headsD · 架构一根主干，三个专家头
The shared torso emits one 2,048-number vector per character. Three heads read them: region & date from the first output; restoration from each masked position. Assael 2022共享主干为每个字符产出 2,048 维向量。三个头读取：地区与年代读第一个输出；修复读每个被遮位置。Assael 2022
🖼 figure: architecture图：architecture
Published architecture (Aeneas variant — adds an image input Ithaca lacks).Published architecture (Aeneas variant — adds an image input Ithaca lacks).
已发表架构（Aeneas 变体 —— 多了 Ithaca 没有的图像输入）。已发表架构（Aeneas 变体 —— 多了 Ithaca 没有的图像输入）。
11 / 27
E · How it learnsE · 如何学习
E · trainingLearning by deliberate damageE · 训练以蓄意损坏来学习
To teach restoration, the team hide parts of intact texts and grade the recovery. This is multitask learning — and the task losses combine with fixed weights. Assael 2022为教会修复，团队把完整文本的部分遮盖，再评分。这是多任务学习 —— 各任务损失以固定权重合并。Assael 2022
▶ interactive: lossStreams — open the live deck to use it交互演示：lossStreams —— 打开实时演示以使用
12 / 27
D · sparse attentionD · 稀疏注意力
D · architectureFull attention vs BigBird: n² → linearD · 架构全注意力 对 BigBird：n² → 线性
Full self-attention compares every pair — ~768×768 ≈ 590,000 per head per layer. BigBird’s sparse attention (global + window + random) cuts that to ~linear; Ithaca stacks 8 blocks. Zaheer 2020全自注意力比较每一对 —— 每头每层约 768×768 ≈ 59 万。BigBird 的稀疏注意力（全局 + 窗口 + 随机）降到近似线性；Ithaca 堆叠 8 块。Zaheer 2020
▶ interactive: sparseattn — open the live deck to use it交互演示：sparseattn —— 打开实时演示以使用
13 / 27
E · The run + codeE · 训练与代码
E · down to the codeThe exact recipeE · 直抵代码精确配方
128128
TPU v4 podsTPU v4 pods
TPU v4 节点TPU v4 节点
~1 wk~1 wk
training timetraining time
训练时长训练时长
8,1928,192
batch sizebatch size
批大小批大小
LAMBLAMB
lr 3×10⁻⁴lr 3×10⁻⁴
Here is the real source code. Toggle “Explain” for a note on each highlighted line. (Aeneas-era repo; differences from Ithaca-2022 flagged.) Assael 2022这里是真实源代码。打开“讲解”查看每个高亮行的注解。（Aeneas 时期的库；与 Ithaca-2022 不同之处已标注。）Assael 2022
▶ interactive: code — open the live deck to use it交互演示：code —— 打开实时演示以使用
14 / 27
E · HonestyE · 诚实
E · a subtle danger“History from square brackets”E · 一个微妙的危险“方括号里的历史”
I.PHI already contains scholars’ own restorations. If the model learns from those guesses, isn’t it echoing them as “fact”? Badian 1989I.PHI 本身已含学者自己的修复。若模型从这些猜测中学习，岂不是把它们当作“事实”回放？Badian 1989
The paper meets this head-on: outputs are framed as plausible suggestions, not certainties — a ranked list, a probability curve, a saliency map. The human stays the judge. Assael 2022论文正面回应：输出被定位为合理建议，而非定论 —— 一份排名、一条概率曲线、一张显著性图。人类始终是裁判。Assael 2022
15 / 27
Lineage · Pythia → Ithaca谱系 · Pythia → Ithaca
LineageWhat Ithaca added to Pythia谱系Ithaca 在 Pythia 之上添了什么
Ithaca’s direct ancestor is Pythia (2019) — the first deep-learning model for ancient-text restoration. Assael 2019Ithaca 的直系祖先是 Pythia（2019）—— 首个用于古文修复的深度学习模型。Assael 2019
Pythia (2019): restoration only · known-length gaps · a single best guess · 30% CER.Pythia（2019）：只做修复 · 缺口长度已知 · 单一最佳猜测 · 30% CER。
Ithaca (2022): + region + date · ranked top-20 · saliency · 62% restoration.Ithaca（2022）：+ 地区 + 年代 · 排名前 20 · 显著性 · 62% 修复。
16 / 27
F · Restoration resultF · 修复结果
F · resultsThe headline: 1 + 1 > 2F · 结果要点：1 + 1 > 2
On 7,811 held-out inscriptions the key result is that human + machine beats either alone — the historian using Ithaca reaches 72%, up from 25%. Assael 2022在 7,811 条留存铭文上，关键是人类 + 机器胜过任何一方 —— 史学家借助 Ithaca 达 72%，较 25% 大增。Assael 2022
▶ interactive: race — open the live deck to use it交互演示：race —— 打开实时演示以使用
17 / 27
F · PlaceF · 地点
F · geographic attributionWhere was it written?F · 地理归属它在哪里书写？
Given only the text, Ithaca ranks the 84 regions — 70.8% top-1, 82.1% top-3, vs 21% for the onomastics baseline. Assael 2022仅凭文本，Ithaca 为 84 个区域排序 —— top-1 70.8%、top-3 82.1%，而人名学基线为 21%。Assael 2022
▶ interactive: regionEx — open the live deck to use it交互演示：regionEx —— 打开实时演示以使用
18 / 27
F · beam searchF · 集束搜索
F · decodingBeam search: many guesses, surest firstF · 解码集束搜索：多猜，最有把握者先
Ithaca keeps a beam of 20 hypotheses and fills the most confident position first — non-autoregressively. Step through εποιη[σε]. Assael 2022Ithaca 保留 20 个假设的束，并先填最有把握的位置 —— 非自回归。逐步走过 εποιη[σε]。Assael 2022
▶ interactive: beam — open the live deck to use it交互演示：beam —— 打开实时演示以使用
19 / 27
F · DateF · 年代
F · chronological attributionWhen? A curve, not a pointF · 年代归属何时？一条曲线，而非一个点
Ithaca outputs a probability distribution over 160 decade-bins — honest about doubt, and within ~29 years on average (median 3). Assael 2022Ithaca 在 160 个十年区间上输出概率分布 —— 对疑问诚实，平均误差约 29 年（中位 3）。Assael 2022
▶ interactive: dating — open the live deck to use it交互演示：dating —— 打开实时演示以使用
20 / 27
G · SaliencyG · 显著性
G · built for trustShowing its work: saliencyG · 为信任而建给出依据：显著性
A saliency map colours the characters that drove a prediction — here “Nikias” and στρατεγοις — so the scholar can check why. Sundararajan 2017显著性图为驱动预测的字符着色 —— 此处是“尼基阿斯”与 στρατεγοις —— 使学者能核查为什么。Sundararajan 2017
▶ interactive: saliency — open the live deck to use it交互演示：saliency —— 打开实时演示以使用
21 / 27
G · Meaning-spaceG · 意义空间
G · contextualisationA map of meaningG · 语境化一张意义的地图
Because Ithaca turns each inscription into an embedding, historically similar texts land near each other — even with no shared words. Assael 2022由于 Ithaca 把每条铭文变成一个嵌入，历史上相似的文本会彼此靠近 —— 即便没有共同词语。Assael 2022
Why it mattersThis “neighbours = similar” geometry beats literal string-matching for finding parallels — and is the exact capability Aeneas turns into contextual retrieval.为何重要这种“相邻即相似”的几何在寻找平行文本上胜过字面匹配 —— 也正是 Aeneas 化为语境检索的能力。
22 / 27
I · Redating AthensI · 重定雅典年代
I · case studyRewriting Athenian historyI · 案例研究改写雅典史
On disputed decrees held out of training, Ithaca’s dates independently align with the modern later readings — avg 5 years from the new ground truths vs 27 for the old labels. Assael 2022在未参与训练的有争议法令上，Ithaca 的定年独立地与现代较晚读法一致 —— 平均距新真值 5 年，旧标注 27 年。Assael 2022
The Chalcis decree (IG I³ 40) → the 420s, not 446/5; the Kleinias decree (IG I³ 34) → 424 BC. An AI can push back against an inherited consensus.卡尔基斯法令（IG I³ 40）→ 公元前 420 年代，而非 446/5；克莱尼阿斯法令（IG I³ 34）→ 前 424 年。AI 能反推承袭已久的共识。
23 / 27
J–K · Limits & impactJ–K · 局限与影响
J–K · honest limits, open toolsWhat it is — and isn’tJ–K · 坦诚局限，开放工具它是什么 —— 又不是什么
Limits: not an oracle — every output needs expert confirmation; Greek only; ≤768 chars; 40% of PHI lacked usable dates.局限：并非神谕 —— 每条输出都需专家确认；仅希腊语；≤768 字符；PHI 中 40% 缺可用日期。
Open: code (github.com/deepmind/ithaca), data pipeline (sommerschield/iphi), a free public interface. The method generalises to any ancient language.开放：代码、数据流程、免费公共界面。该方法可推广到任何古代语言。
24 / 27
L · → AeneasL · → Aeneas
L · the lineage continuesL · the lineage continues
L · 谱系延续L · 谱系延续
Pythia → Ithaca → AeneasPythia → Ithaca → Aeneas
In 2025 the same team released Aeneas: it adds Latin, reads the inscription’s image, restores gaps of unknown length, and retrieves contextual parallels.2025 年，同一团队发布了 Aeneas：它加入拉丁语、读取铭文图像、修复未知长度缺口，并检索语境平行文本。
Open the Aeneas deck →Open the Aeneas deck →
打开 Aeneas 演示 →打开 Aeneas 演示 →
The lineage →The lineage →
技术谱系 →技术谱系 →
25 / 27
M · From experiment to toolM · 从实验到工具
M · from experiment to better toolWhat this means for SymbolonM · 从实验到更好的工具这对 Symbolon 意味着什么
Symbolon — your layer that orchestrates Ithaca + Aeneas — is one experiment in human-AI epigraphy, not a finished paradigm.Symbolon —— 你编排 Ithaca + Aeneas 的那一层 —— 是人机金石学的一次实验，而非定型范式。
Keep: ranked hypotheses; dating as a distribution; saliency surfaced; the human as judge.保留：排名假设；定年作为分布；呈现显著性；人类作裁判。
Improve: surface saliency in notes; preserve the full date curve; bridge Greek↔Latin at the concept layer; use Aeneas’ image head.改进：在笔记中呈现显著性；保留完整定年曲线；在概念层桥接希腊语↔拉丁语；用 Aeneas 图像头。
DeliverableWritten up as SYMBOLON_IMPROVEMENTS.md — each tied to a paper passage and a code site.交付物整理为 SYMBOLON_IMPROVEMENTS.md —— 每条对应一段论文与一处代码。
26 / 27
Sources & reading出处与延伸阅读
SourcesEvery claim, traceable出处每个论断，皆可溯源
▍ local source · ▍ web source. The companion Lineage deck threads these into the full technological story.▍ 本地来源 · ▍ 网络来源。配套的谱系演示把它们串成完整的技术故事。
▶ interactive: sources — open the live deck to use it交互演示：sources —— 打开实时演示以使用
27 / 27
Thank you谢谢
From square bracketsto a partner从方括号到一位伙伴
Ithaca did not replace the historian. It gave them a faster, humbler, more transparent collaborator — one that shows its work and hands back the final word.Ithaca 没有取代史学家。它给了他们一位更快、更谦逊、更透明的合作者 —— 一个会展示依据、并把最终决定权交还的伙伴。
Continue to Aeneas →Continue to Aeneas →
继续看 Aeneas →继续看 Aeneas →
Symbolon →Symbolon →
Symbolon →Symbolon →
The full journey →The full journey →
完整旅程 →完整旅程 →
Assael, Sommerschield et al., Nature 603, 280–289 (2022) · doi 10.1038/s41586-022-04448-zAssael, Sommerschield et al., Nature 603, 280–289 (2022) · doi 10.1038/s41586-022-04448-z
© 2026 Wu Ching-Yuan 吴靖远 · magalia.wiki (籬廬). Generated transcript 2026-06-13 from ithaca.html · text CC BY 4.0. Papers © their authors (DeepMind, Nature).