EpiDoc = TEI subset for inscriptions, optimised for interchange EpiDoc = 面向铭文的 TEI 子集，专为交换而设

EpiDoc is not a new schema. It is a specialised set of guidelines for using the Text Encoding Initiative (TEI) XML — the same XML standard used in the literary, linguistic, and manuscript communities. Building on TEI means maximum compatibility with the wider digital-humanities ecosystem.

The architectural separation matters: the data (what the inscription is, what was lost, what was restored, who edited it, when, on what authority) is separate from the view (HTML for the web, PDF for print, audio for the visually impaired, a database row for aggregation). One XML source supports many views; flat-text publication couples the two and discards everything that wasn't needed for the printed page.

One inscription, three ways to read it 一篇铭文的三种读法

Bodard's worked example is the Aphrodisias text ALA 2 — a fragmentary statue base honouring Salonina (Julia Cornelia Salonina Augusta). The text below is the one published in Aphrodisias in Late Antiquity (Roueché 1989, ALA 2).

The same data, additional layers

Once a word is in the XML, additional information can be attached non-destructively. Two examples from Bodard's discussion:

Real editorial chain · ISic000118 over 9 years

Bodard argues that EpiDoc captures information no print edition could carry: the full editorial history of a record, with named contributors and date stamps. Here is the actual <revisionDesc> from ISic000118.xml in the mounted demonstrator — a single funerary epitaph whose digital edition has been touched 11 times by 6 named editors over 9 years:

Each who="#JCu" reference resolves to a <respStmt> earlier in the file. Some of those entries carry ORCID identifiers — Prag's http://orcid.org/0000-0003-3819-8537, Stoyanova's 0000-0003-3914-9569, Cummings's 0000-0002-6686-3728, Ahlholm's 0000-0001-8417-7089, Crellin's 0000-0002-0100-7437, the petrographer Coccato's 0000-0002-6641-2820 — making the editorial history not just citable but globally addressable to real persistent researcher identities.

One source, many outputs — via XSLT 单一源数据，多种产出，经由 XSLT

The capability that makes structured preservation worth the cost: multiple deliverables from one source, generated automatically. A single XSL Transformation script can:

• HTML for the web — Leiden-style transcription with interactive footnotes
• PDF for typeset print — print-on-demand volume from the same data
• Diplomatic version — strip out editorial restorations to show what is actually on the stone
• Indexed onomasticon — automatically generated from <name> elements across the corpus
• Concordance — every attestation of every <w lemma="..."> grouped by lemma
• Audio — for visually impaired readers, with editorial conventions read aloud
• Database row — for aggregation into EDH, EDR, EDCS or Wikidata
• Translation alignment — line-by-line parallel text in any other language

The corollary: every transformation is reversible. If a new question requires a different output, the XSLT is rewritten, the same XML is re-transformed, and the new output appears. The data has not been touched. The data does not need to be touched, because the data already preserves the structure the new question depends on.

What a project needs · what a scholar needs to learn 项目所需 · 学者需学的技能

Bodard is honest about the cost. EpiDoc requires more skill than typing Leiden into Word. But the skills are not technical novelties — they are editorial decisions made explicit.

• Scholar: learns to read EpiDoc XML well enough to review work; understands what each tag means as an editorial distinction; engages with the EpiDoc community via lists and fora.
• Programmer / RA: customises XSLT for the project; adapts CHET-C or similar Leiden→XML converter; ties the output into a web design.
• Web designer: wraps the XSLT-generated HTML into a coherent site.
• Not needed: a proprietary database system that locks the data away; reinventing tag conventions; software that depends on a single vendor.

Take-away for Week 12 · 本周要点

Bodard's paper is the practical companion to Cayless's vision. Cayless asks why preserve; Bodard answers how. When the student in Week 12 writes their first <supplied reason="lost">, they are doing exactly what Bodard's paper describes — turning editorial judgement into machine-readable structure. The verbosity is the visible cost; the reusability is the invisible benefit, paid back many times over as the same data is rendered, indexed, aggregated, translated, and queried.

An inscription has more than text · 一篇铭文不止是文本

Sicily across 1,500 years carries Greek, Latin, Punic, Elymian, Sikel, and Oscan writing on stone, ceramic, brick, metal, and wax. The Crossreads project asks what happens when every dimension of these inscriptions — not just the text — is preserved as linked structured data.

Connecting letterforms to text, material, date, place 将字形与文本、材质、年代、地点相互连接

Stoyanova & Prag describe what they call the "main challenge" — designing an effective mechanism to integrate information from multiple datasets to enable rich and complex queries. The integration is achievable only because each layer was, individually, preserved as structured data with stable identifiers and linked-open-data anchors.

An example from a real I.Sicily file (ISic000118, a funerary epitaph from Thermae Himeraeae) shows what structured preservation of just one dimension — material — looks like:

This single element preserves: the material name in English (conglomerate); a taxonomy anchor (#material.inorganic.stone); a typological classification (type, subtype); the petrographic specialist who identified it (Coccato, with ORCID elsewhere in the file); and a URI to the EAGLE vocabulary entry that aligns this term with every other epigraphic project's controlled vocabulary for materials.

Crossreads extends this pattern to letterforms

Each letterform observation — alpha with broken bar, alpha with straight bar, sigma four-bar versus lunate — is recorded against the IIIF image region, attributed to the palaeographer who made the observation, anchored to the EpiDoc text-position via stand-off markup, queryable against the material, date, place, and language of the inscription. The palaeographer's judgement becomes data the way the editor's restoration became data in Bodard's example.

Every research dimension is a candidate for structured preservation 每一研究维度都可成为结构化保存对象

The Crossreads project demonstrates a pattern that runs through all six papers in this lecture: any aspect of an inscription that a scholar observes carefully enough to publish, is also an aspect that — if preserved as structured data — supports computational analysis at scale.

Real evidence · different corpora preserve different things

The rubric scored at population scale (rubric_full.json, 10,249 source files across all four corpora) shows a striking pattern: encoding traditions specialise. Each corpus is strong on different rubric axes, exactly as Stoyanova & Prag predict for a multilingual, multi-period, multi-material region like Sicily — where no single encoding decision is universally optimal.

Take-away for Week 12 · 本周要点

EpiDoc is not only for the inscription's text. It is for everything the scholar observes about the inscription — including aspects (letterforms, petrography, editorial hands) that traditional print could not even document well. The XML structure is what makes those observations linkable to each other, and therefore queryable, and therefore actually usable for the kinds of large-scale historical and linguistic questions the field wants to ask.

What if your inscription is in Faliscan, not Latin? 如果你的铭文不是拉丁文，而是法利斯语呢？

The ItAnt project (Italian PRIN, CNR-ILC Pisa + Ca' Foscari Venice + University of Florence) focuses on the languages of pre-Roman Italy attested only in epigraphic form — Restsprachen, "remnant languages." Their epigraphic record is fragmentary, often uncertain in reading and segmentation, and presents challenges the standard EpiDoc tagset (built first for Greek and Latin) does not directly handle.

Murano et al.'s response is not to invent a new schema. It is to extend EpiDoc with carefully chosen additions that record exactly the distinctions these materials require — preserving them as structured data, while keeping interoperability with the international EpiDoc community.

The ItAnt EpiDoc extensions ItAnt 的五项 EpiDoc 扩展

What flat-text would discard 平面文本会丢弃什么

Consider an Oscan inscription written in Greek script. In a flat-text edition you would see the Greek letters and a footnote saying "Oscan." In EpiDoc with ItAnt's customisations:

That XML says, machine-readably:

• The language is Oscan (ident="osc") — searchable, aggregateable, joinable to Wikidata and linguistic resources.
• The script is Greek (in its Oscan-adapted form) — independently queryable. "Show me all Oscan inscriptions written in Greek script" is one query; "show me all Oscan inscriptions written in Oscan script" is another. Both are answerable.
• Words are separated by punctuation, not by spaces — so any text-processing tool knows how to tokenise this particular inscription correctly.

Real example · how fully-marked-up Greek looks at scale

The principle Murano et al. apply to fragmentary Italian languages applies equally to canonical Greek and Latin. Here is a fragment from A.30.xml (Anastasius I's imperial edict at Apollonia, c. 491–518 CE) in the mounted IRCyr corpus — every word marked with <w lemma="...">, every name marked with <persName> + <name> + nymRef, every abbreviation marked with <expan> + <abbr> + <ex>, every lost letter marked with <supplied reason="lost">:

Counting only Greek-script <w lemma=> elements across the four EpiDoc inscription corpora, the demonstrator's federation lemma resource federation_lemmas_full.json records 41,411 attestations. That is the linguistic substrate ItAnt-style customisations enable at federation scale. Each token-lemma pair preserves: surface form, normalised surface, lemma, normalised lemma, line number, file, corpus, language, provenance, licence. No print edition could carry this density.

Beyond EpiDoc: connecting to the cultural-heritage semantic web 超越 EpiDoc：接入文化遗产语义网

ItAnt does not stop at extending EpiDoc. Murano et al. also encode the inscriptions against the CIDOC CRM conceptual reference model (the ISO 21127 ontology used across European cultural-heritage institutions) and its extensions CRMtex (for ancient texts) and CRMinf (for scholarly arguments and inferences). And bibliographic metadata is encoded in FRBRoo / LRMoo.

The result is a record that is simultaneously: an EpiDoc inscription, a CIDOC-CRM-typed cultural object, a CRMtex-typed text-bearing artefact, a CRMinf-typed scholarly argument, and an LRMoo-typed bibliographic citation. Each lens supports different queries. The data exists once; the queries — and the interoperability with non-epigraphic cultural-heritage databases — multiply.

Take-away for Week 12 · 本周要点

Murano et al.'s paper shows that EpiDoc is not a monolithic standard you accept or reject. It is a foundation that extends to your specific research needs. The lesson for the student is twofold: first, every customisation must be principled (use <rs> with controlled vocabularies, not invent new elements); second, every customisation should serve preservation — adding categorical distinctions the standard does not yet carry, never collapsing distinctions the standard already records.

EpiDoc XML preserves everything — and is painful to write EpiDoc XML 保留一切，也极难手写

The cost of preservation is verbosity. Each <supplied reason="lost"> is twenty-four characters of tag for one character of editorial decision. Each opening tag matches a closing tag. Each attribute name and value must be written in full. The percentage of information content to structural scaffolding is unbalanced — and human readability decreases rapidly as complexity increases.

In ItAnt, the data density is high: linguistic, philological, and prosopographical information overlaps. A lacuna spans the end of token 3 and the start of token 4. A named entity (praenomen partially conjectured, gentilicium, patronymic) spans tokens 4 through 6. These overlapping hierarchies are the well-known weakness of XML, addressed by stand-off markup or alternative representations — but both add complexity for the human encoder.

Boschetti et al. test this with a real philologist (linguistics graduate, epigraphic competence, basic DH skills only). She encodes five Faliscan inscriptions twice — once in raw EpiDoc, once in ItAntDSL. The results are qualitative but consistent: training time on the DSL is significantly shorter; encoding time is markedly lower; DSL documents are approximately three times more compact than the EpiDoc counterparts.

Concrete · how big is one EpiDoc file, really?

Look at the actual ISic000118.xml in the mounted demonstrator. This is one funerary inscription. Four lines of Latin on the stone. What is preserved in XML, measured in bytes:

How the DSL becomes EpiDoc XML DSL 如何转译为 EpiDoc XML

Key design choices that make this work:

• ANTLR for parsing. The DSL has a formal context-free grammar (available on GitHub at CoPhi/itantdsl). The parser produces an Abstract Syntax Tree from which an intermediate XML representation is generated.
• Six design dimensions optimised (Zenzaro et al. 2022): familiarity, transparency, completeness, compactness, consistency, actionability. The DSL wins on the first three; XML wins on actionability; both target completeness and consistency.
• Shared metadata in YAML. Information repeated across every inscription (language definitions, script definitions, vocabulary aliases) lives once in a YAML look-up file. The DSL references it by key. The XSLT expansion merges it in. Each inscription stays compact.
• EpiDoc is the canonical artefact. The DSL is the editorial surface. The XML is the data that ships. ItAnt does not invent an alternative standard — the international ecosystem still receives standard EpiDoc.

Ergonomics serves preservation — does not bypass it 人机工效服务于保留，而非绕开保留

The temptation, when faced with EpiDoc's verbosity, is to say "I'll just use a flat text and skip the XML." That is the move Boschetti et al. explicitly reject. The DSL is faster and more compact and easier to learn, but it produces the same full EpiDoc XML the field expects. The structure is preserved end-to-end.

Connection to your Week 12 hands-on · 与本周实践的关系

When you first try to write inscription data in EpiDoc, you will feel the friction Boschetti et al. document. That is normal. The friction is a sign that you are being asked to articulate distinctions that flat-text writing simply did not require you to make. The friction is the cost of preservation.

For your first attempts, write the EpiDoc by hand. Feel where it pinches. Notice which decisions feel laborious (line breaks? abbreviations? names?). Then — and only then — would it be sensible to look for or build a DSL or template editor for your specific corpus. The DSL is a useful tool, but the EpiDoc literacy is the foundational skill, and the friction of writing it raw is part of how you build that literacy.

From CIG (1815) to FAIR Epigraphy (2022): why corpora exist 从 CIG (1815) 到 FAIR 铭文学 (2022)：语料库为何存在

Boeckh's 1815 proposal to the Berlin Academy already had the core elements: comprehensive geographic coverage, autopsy-based editions, indices of names. What he could not have anticipated: that the 19th-century print model would, two centuries later, encounter limits that digital methods now have to address.

Bodel/Prag/Roueché's claim: digital methods, in particular EpiDoc TEI for encoding and Linked Open Data for connection, enable epigraphic corpora to capture far more than Leiden or its codex-bound print conventions could ever carry. The question is no longer whether digital is possible — it is whether the community is willing to adapt its practice.

Why "CIL I² 1221 = VI 9499 = ILS 7472 = CLE 959 = ILLRP 793" is wrong 为何"等号串联多版编号"的引用方式是错误的

One of the paper's sharpest critiques targets a practice almost every epigraphist has engaged in: citing an inscription by stringing together several corpus numbers separated by equals signs, as if they were identifiers for the same text rather than separate editions by separate scholars at separate times.

The citation chain conceals authorship. It treats edition as if it were identifier. And in the digital age this practice is, if anything, more dangerous. The EDCS entry might be a copy of an existing edition, or it might be modified — without a date, without a named editor, without a documented method. The PHI text might or might not reflect a particular published edition. The user has no way to know.

What structured preservation enables instead

The contrast is stark. EDR's edition is structured, attributed, and citable to a particular scholar at a particular time, because the EpiDoc XML preserves those things. EDCS's entry — even though it exists in the same digital medium — is opaque because the underlying data structure did not preserve provenance. The opacity is not because the medium is digital; it is because the data was not structured to preserve authorship and method.

From IRT (1952) → IRT2009 → IRT2021 — preservation enabled iteration 从 IRT 印本 (1952) 到数字版 IRT2021，保留促成迭代

Joyce Reynolds and J.B. Ward-Perkins published The Inscriptions of Roman Tripolitania in 1952, post-war, with limited photography (38 plates for 1000 texts). Louis Robert promptly noted the limitations. The volume served the field for half a century.

In 2009 the British School at Rome, ISAW (NYU), and KCL collaborated to publish IRT2009 online — the 1952 text re-encoded in EpiDoc XML with images, translations, and structured metadata. Because the encoding preserved structure, when scholars wanted updates, those updates could be applied surgically. Joyce Reynolds provided English translations. Place identifications got Pleiades URIs. Editorial decisions stayed traceable.

In 2021, the same data was enriched: every Greek or Latin inscription from Tripolitania published since 1952 was added. The result, IRT2021, took less than 12 months to produce. Standard EpiDoc encoding meant the existing data did not have to be re-keyed. New inscriptions slotted in. Indices auto-regenerated. Wikidata links to LGPN (Lexicon of Greek Personal Names) and PIR (Prosopographia Imperii Romani) propagated.

This is what Cayless's 2009 vision of "true datasets able to be queried, mined, and transformed" looks like in practice, fifteen years later. The corpus is no longer a static publication. It is a living, citable, versionable dataset, and the cost of updating it is bounded by the cost of writing new editions — not by the cost of resetting the entire print.

The future of the corpus: dynamic, federated, FAIR 语料库的未来：动态的、联邦的、FAIR 的

Bodel/Prag/Roueché's closing proposal is an inversion of the existing corpus model: instead of a single monumental corpus (CIG, CIL, IG) covering everything, build directly on the proliferation of local and thematic corpora (IRT, IRCyr, I.Sicily, USEP, PETRAE, etc.) and connect them via standards.

The principles · FAIR data

• Findable: unique persistent identifier (DOI, TM, URI), rich metadata in a searchable resource
• Accessible: retrievable by its identifier, using open/free protocols, with metadata accessible even when the data is restricted
• Interoperable: shared formal vocabularies (EAGLE, FAIR Epigraphic Vocabularies, Getty AAT), language-aware, machine-actionable
• Reusable: clear licensing, accurate provenance, community standards, rich documentation

The pilot · inscriptiones.org

The FAIR Epigraphy project (AHRC + DFG, dir. Horster & Prag) has demonstrated the pattern. A software engineer with no epigraphic background, working a few hours, transformed XML files from Roman Inscriptions of Britain (RIB), Inscriptions of Greek Cyrenaica (IGCyr), and I.Sicily into RDF using a community-built ontology. The output is queryable at https://inscriptiones.org/.

What this looks like in the mounted demonstrator

The same demonstrator that hosts these slides ships its own Pelagios TTL serialisation at /data/json/federation_pelagios.ttl, exporting all 56 records as oa:Annotation objects linked to pleiades: places. This is the actual output format Bodel/Prag/Roueché describe — a real RDF file, openly downloadable, sitting at federation scale:

Each inscription becomes a tuple in the Pelagios graph. Each places-link makes the inscription visible to Peripleo and any other Pelagios consumer. The full federation TTL is 336 triples across 56 records — a small file, but a working contribution to the international LOD graph that aggregators like inscriptiones.org and the Trismegistos cross-reference graph can ingest.

Real proof · the international encoding tradition is internally consistent

One of Bodel/Prag/Roueché's underlying assumptions — that structured preservation across corpora is meaningful enough to aggregate — is testable from the same demonstrator data. The cross-encoder consistency analysis (cross_encoder_full.json) measures: across 41,411 hand-attested Greek lemma instances from four EpiDoc inscription corpora, how often do different editorial teams give the same Greek surface form the same lemma?

What the community has to do · 学术共同体应做的事

• Cite digital editions properly: author, date, version, URL
• Treat each edition as a scholar's editorial work — not as an opaque database row
• Agree on community standards for categories (inscription type, material, support) — leaving how the categories are represented free
• Document the construction of editions: who recorded the metadata, who took the photograph, who compiled the bibliography
• Use ORCIDs to track contributions in collaborative environments
• Adopt versioning (Zenodo deposits, GitHub history, semantic versioning of corpus editions)
• License data openly (CC-BY-4.0 is the field's emerging default)
• Stop citing inscriptions by equals-sign chains that obscure authorship
• Stop publishing data in HTML-only interfaces with no bulk export

Take-away for Week 12 · 本周要点

When you write inscription data in EpiDoc for the first time this week, you are participating in the system Bodel/Prag/Roueché describe. Your editorial decisions become part of a federation that — if the community standards hold — can be cited, versioned, aggregated, queried, and re-used by anyone in the field, forever. The cost of preservation is real. The dividend is the world the three authors are arguing for: a research culture in which our data is as open, citable, and re-usable as our arguments have always claimed to be.

What the six papers say together 六篇论文的合鸣

Now — your hands-on · 现在，动手实践

Try writing inscription data in EpiDoc. Pick a short inscription from the mounted /Users/chingyuanwu/Documents/epidoc/isicily/m0-demonstrator/data/xml/ corpus — perhaps ISic000118 (Latin epitaph) or A.30 (Greek imperial edict from Cyrene) — open its existing EpiDoc XML, and read it line by line. Then try writing your own short EpiDoc record from scratch for a Greek or Latin text of your choosing. Notice what the standard tagset gives you. Notice what your specific corpus might require beyond the standard. Notice which editorial decisions you are now recording as data that you might previously have left in your head or in a footnote.

The data you write this week is preservation infrastructure. Future you, and future others, will be able to ask it questions you have not yet thought of.

Mounted folders for hands-on work: /Users/chingyuanwu/Documents/epidoc/isicily/ · /Users/chingyuanwu/Documents/epidoc/kcl_tei/ (Aphrodisias, Cyrenaica, Tripolitania) · /Users/chingyuanwu/Documents/epidoc/inscription_databases/. The companion Leiden ⇄ EpiDoc playground is shipped at /Users/chingyuanwu/Documents/epidoc/isicily/m0-demonstrator/leiden-playground.html.

Concrete artefacts to open today 今日可亲手打开的具体文件

Every claim in this deck was anchored in real files. The mounted /Users/chingyuanwu/Documents/epidoc/ folder hosts all of them. Here is the suggested walkthrough order: