Intertextuality in diachronic Greek

What AthDGC contributes; Symposium on Computational Methods for Intertextuality, Tel Aviv University, 22 June 2026

Nikolaos Lavidas

NKUA · Athens Digital Glossa Chronos Research Network

June 22, 2026

Intertextuality in diachronic Greek

What AthDGC contributes

Nikolaos Lavidas, NKUA Athens Digital Glossa Chronos Research Network Διαχρονία Γλώσσας :Χρόνος

HFRI Project No. 20577 · Greece 2.0 NRRP · Compute GRNET ARIS pa260305

Good morning. Thank you to Nachum, Kfir, and Kai for the invitation, and to the RUNI-TAU group for putting this symposium together. I am Nikolaos Lavidas from the National and Kapodistrian University of Athens, and I will speak today about diachronic Greek, about a particular hypothesis on written language contact, and about the AthDGC platform that operationalises both.

Roadpath

1. The premise. Historical linguistics works on texts, not speakers.
2. What intertextuality means inside a diachronic-Greek setting.
3. Three modes: retelling, retranslation, inner-textual citation.
4. The hypothesis: parallel coexistent grammars in written contact (Lavidas 2021, Brill).
5. How AthDGC operationalises the hypothesis: one PROIEL schema across 3000 years.
6. Concrete platform, concrete numbers, three threads through the corpus.
7. What this offers a computational-intertextuality programme.

Seven stops. I will spend most of the time on stops 3, 4, 5, 6. Stops 1, 2, 7 frame the rest.

The AthDGC team

Nikolaos Lavidas (PI, NKUA) · Kiki Nikiforidou (NKUA) Dag Haug (Oslo, PROIEL Director) · Leonid Kulikov (Ghent) Vassiliki Geka (NKUA) · Vassileios Symeonidis (NKUA) · Theodoros Michalareas (NKUA) Sofia Chionidi (NKUA) · Anastasia Tsiropina (NKUA) · Eleni Plakoutsi (NKUA) · Evangelos Argyropoulos (NKUA)

Three universities (NKUA · Oslo · Ghent); eleven researchers; one platform. HFRI Project No. 20577 · Greece 2.0 NRRP.

AthDGC is the Athens node of the PROIEL family (“Athens-PROIEL”). Dag Haug, founding director of PROIEL at Oslo, is a co-author and co-PI. We adopt the PROIEL XML 2.0 schema verbatim and we extend it diachronically.

Eleven authors, three universities. The PROIEL group at Oslo under Dag Haug and the Ghent comparative IE group under Leonid Kulikov are the international partners; the NKUA core is led by Kiki Nikiforidou and me, with three post-doctoral researchers (Vassiliki Geka, Vassileios Symeonidis, Theodoros Michalareas), three doctoral candidates (Sofia Chionidi, Anastasia Tsiropina, Eleni Plakoutsi), and one research assistant (Evangelos Argyropoulos).

1. The premise

Historical linguistics has no native speakers. It has only texts.

The only window onto Archaic, Classical, Koine, Late Antique, Byzantine, Late Byzantine, and Early Modern Greek is the written record. Every claim about the syntax of a period reduces to a claim about what its surviving texts permit and what they exclude.

This is the bedrock for everything that follows. We cannot run an acceptability survey on Homer. We cannot ask a fluent Byzantine to judge whether a sentence is well-formed. Whatever we can say about the grammar of those periods, we have to recover from what writers wrote, and from how their text was copied, cited, retold, and retranslated by writers in later periods. That last sentence is already an intertextuality claim, and it is the bridge to today’s theme.

2. Intertextuality in a diachronic-Greek setting

In a corpus that spans 3000 years on one language, intertextuality is not a stylistic ornament. It is the primary data-generating process.

The same canon is re-rendered by every generation:

• the same Homeric verse is paraphrased by Byzantine schoolbooks;
• the same Pauline epistle is retranslated into Modern Greek every century;
• the same Septuagint formula is quoted, glossed, and inflected toward each period’s spoken grammar.

For Hebrew studies, the parallel is exact: Mishnah quotes Tanakh; Talmud quotes Mishnah; mediaeval commentary quotes Talmud. For Buddhist Sanskrit and Tibetan, the same: a single sutra is rendered into Tibetan, into Chinese, into Mongolian, with each rendering registering the syntax of its receiving tradition. The pattern is universal. Greek just happens to have the longest continuous documented chain in the Indo-European family, and that is what makes it a useful laboratory.

3. Three modes

Retelling. Same language, later century. Same story, new register. (Niketas Choniates retelling Homer in 12th c. Byzantine epitome.)

Retranslation. Same canon, target language shifts. (Hebrew Bible → Greek Septuagint → Latin Vulgate → Gothic Wulfila → OCS Marianus → Classical Armenian.)

Inner-textual citation. Surface re-use of one text inside another. (NT quoting LXX; Patristic homily quoting NT; Byzantine chronicle quoting Patristic.)

Each of the three modes leaves a recoverable trace in the syntax. Retelling re-encodes the same propositional content in the grammar of a later period. Retranslation re-encodes the same propositional content in the grammar of a sister Indo-European language. Inner-textual citation embeds a frozen older clause into a host text whose grammar is younger. All three are queryable, if the corpus is annotated uniformly.

4. The hypothesis

Under sustained written language contact, donor and recipient grammars do not displace one another. They coexist as distinct active systems inside the same writer.

Lavidas, N. 2021. The Diachrony of Written Language Contact: A Contrastive Approach. Brill’s Studies in Historical Linguistics. Leiden, Boston: Brill.

This is the central claim of the 2021 monograph and the lens through which AthDGC reads its own data. The hypothesis is that in oral contact one of the two grammars typically wins out within a generation or two. In written contact, the donor grammar persists alongside the recipient grammar for centuries, because written transmission has a copy-faithful mechanism that oral transmission lacks. The writer of a Byzantine homily on Paul is not choosing between Atticist syntax and his own vernacular: both are active, both are accessible, and the choice between them is register-driven, not displacement-driven.

5. Why “written” matters

Written transmission has its own physics:

• prestige standards (Atticism, Byzantine Hochsprache);
• copying chains with attested manuscript stemmata;
• citation as authority (citing scripture as scripture);
• resistance to vernacular pressure at literate registers.

Each of these leaves a different signature in the syntactic record.

For an intertextuality programme, this is good news: the “physics” of written transmission is precisely what intertextuality detection systems already exploit. Copy-faithful mechanisms produce surface matches; prestige mechanisms produce learned-vocabulary clusters; citation mechanisms produce the explicit boundaries that probabilistic text-reuse detectors are tuned for. Diachronic syntax simply adds the structural dimension: the same propositional content, re-encoded into a syntactically different host.

6. One schema across 3000 years

PROIEL XML 2.0 is the file format that stores each sentence as a tree of word-by-word grammatical relations. Developed at Oslo for the early Indo-European languages; extended in AthDGC to all of Greek.

The same 26-relation inventory annotates:

• Homer (8th c. BCE)
• the New Testament (1st c. CE)
• Sphrantzes’ Chronicle (15th c. CE)
• the Modern Greek standard (21st c. CE)

One schema; eight periods; queryable comparison.

The diachronic extension was first proposed by Lavidas and Haug (2012, Thessaloniki-Oslo PROIEL pilot on Sphrantzes’ Chronicle). AthDGC v0.4 ships that 2012 idea, generalised to the full diachronic span.

The PROIEL group at Oslo, under Dag Haug, established the schema in 2008. The 26 relations cover predication, subjects, objects, obliques, attributes, adverbials, vocatives, auxiliaries, all the coordination and apposition patterns, and the typological relations specific to early Indo-European such as ag (agent of a passive) and xobj (raised object of a control verb). The diachronic extension of that schema, that is the proposal to take the same inventory beyond Koine and Classical into Byzantine and later, goes back to the 2012 Thessaloniki-Oslo PROIEL pilot in which Dag Haug and I annotated a sample of Sphrantzes’ Chronicle (15th c. Late Byzantine) in PROIEL XML. The AthDGC v0.4 release ships that 2012 idea at full scale: same inventory, eight periods, 22-tag source-metadata header.

7. Text Id

Every partition carries a stable identity:

athdgc.<author>.<work>.<src_lang>.<tgt_lang>.<translator>.v<revision>

Example:

athdgc.017.001.grc.000.000.v1 (Sphrantzes, Chronicle, Greek)

Every Text Id resolves to a row in a metadata register, to a source edition, and to a citable printed reference.

For intertextuality work this is the part that pays off most. If you find a verse echo between two AthDGC partitions, the Text Id at each end gives you the edition, the printed citation, the editor, the digitisation, and the manuscript reference, all in one identifier. The metadata register sits behind the Text Id; it is the bridge between the computational object and the philological record.

8. What is built (tools)

• An annotation pipeline that reads a Greek sentence and emits PROIEL XML 2.0 (the file format that stores each sentence as a tree of word-by-word grammatical relations, developed at Oslo).
• A morphological tagger (Stanford Stanza, in its PROIEL-trained version), which produces lemma + part-of-speech + features automatically.
• An AI second-opinion review loop: uncertain annotations are flagged and re-checked by a large language model, then approved by a human.
• A live dashboard that shows token counts and per-period progress in real time as the annotation runs.

All open source. MIT · Apache 2.0 · BSD.

Four pieces, in plain words. The annotation pipeline is the core: Greek sentence in, PROIEL XML tree out. The morphological tagger is Stanza, which is Stanford’s open Python toolkit for natural language processing; we use its PROIEL-trained Greek model. The AI review loop is the part that lets the project run at marathon scale without losing quality: every uncertain call is flagged, sent to a large language model (Claude Haiku, currently) for a second opinion, and the human annotator only has to approve, reject, or amend. The dashboard reads the live ARIS state every three minutes; I will not screen-share it, but it is the picture behind the numbers on slide 11.

9. What is built (corpus)

Eight diachronic periods, all on one PROIEL XML 2.0 spine:

	Period	Status
	Archaic, Classical, Hellenistic, Koine	partitions in flight
	Late Antique, Byzantine	partitions in flight
	Late Byzantine, Early Modern, Modern	partitions in flight

New Testament verse-level alignment (cross-alignment: machine-readable matching of corresponding sentences and words between texts in different languages):

• to Latin (Vulgate), Gothic (Wulfila), Old Church Slavonic (Marianus): aligned in v0.4;
• to Classical Armenian: workflow established, alignment in ingestion for v0.5.

“In flight” means the marathon is actively producing them on ARIS, this week. The samples on the public site are real; the full audited cuts release at v0.5.0. For Classical Armenian: the workflow is in place but the alignment itself is still being built; that lands in the v0.5 cut.

10. What is built (platform)

• Public showcase: athdgc.github.io
• Corpus repository: github.com/AthDGC/athdgc-corpus (closed during v0.5 audit)
• Live PROIEL viewer: dialing.enl.uoa.gr/proiel/
• Launch report: arXiv cs.CL (published 11 Jun 2026)
• Permanent identifier: Zenodo concept DOI 10.5281/zenodo.20439182

CC BY 4.0 for the corpus. Citable per-version DOIs on each release.

The public site, the GitHub organisation, the PROIEL viewer at NKUA, the arXiv preprint, and the Zenodo deposition are all live; you can pull any one of them up while we are speaking.

11. Concrete numbers (today)

10.91 M tokens PROIEL-annotated 173 annotation batches complete 8 Greek periods covered 3 Indo-European parallels currently aligned at the New Testament

Dashboard snapshot: 19 June 2026 21:17 Athens. Latin, Gothic, Old Church Slavonic aligned in v0.4; Classical Armenian in ingestion for v0.5. v0.5 target 16 M tokens; v0.6 target 24 M.

At this size AthDGC is already the largest annotated diachronic Greek corpus in existence (Diorisis at 10.2 M is the previous reference point), the largest PROIEL-style dependency-parsed corpus of any language, and the only diachronic corpus that covers a 3000-year continuous span on one schema. The v0.5 cut takes the lead from 0.7 M to ~6 M tokens.

12. Three threads through the corpus

Thread A. The Homeric chain. Iliad 1.1 in Archaic Greek → its Byzantine epitome (Tzetzes) → its 1955 modern Greek prose retelling (Kakridis-Kazantzakis).

Thread B. The New-Testament chain. John 1.1 in Koine → its Vulgate translation → its Gothic and Old Church Slavonic sisters.

Thread C. The Septuagint chain. Psalm 1.1 in Hebrew (target) ← LXX Greek → Vulgate → Modern Greek liturgy.

Each thread is a real chain of real partitions. A and C are useful for retelling-and-retranslation; B is useful for cross-lingual IE structural comparison. Together they cover the three modes I listed earlier.

12a. A closer look at the Septuagint

The Septuagint (LXX) is the most consequential single text in the Greek chain:

• a Koine-Greek translation of the Hebrew Bible, made in stages from the 3rd c. BCE (Pentateuch) onward;
• the Greek Bible of the early Church: New Testament writers quote it, not the Hebrew;
• the source of the first Christian translations into Latin (Old Latin, then Vulgate for many books), Gothic (Wulfila), Old Church Slavonic (Cyril and Methodius), and Classical Armenian;
• still in use in the Modern Greek liturgy today.

This is the slide that should land most directly with the TAU and RUNI audience. The Septuagint is a hinge: a translation that becomes a source, and a translation that carries Hebrew syntactic structure into the entire later Christian translation tradition. In AthDGC it sits inside the Koine partition, and it is the spine of Thread C.

12b. What LXX shows on a PROIEL spine

Two well-studied phenomena that AthDGC makes queryable across periods:

1. Paratactic καί as a calque of the Hebrew waw-consecutive: clause-initial καί + finite verb, instead of classical Greek subordinate-clause patterns;
2. Closest-conjunct agreement: number and gender agreement track the nearest conjunct under Hebrew influence (Lavidas 2019, Questions and Answers in Linguistics 5(2): 37-90).

Both are syntactic transfers from Hebrew. Both propagate outward, through Vulgate, Gothic, Old Church Slavonic.

These are two phenomena I can speak to with confidence because the second has my own published account behind it (Lavidas 2019). The first, paratactic καί as a calque of the waw-consecutive, is a standard finding in Septuagint Greek studies (Conybeare and Stock, Thackeray, Aitken). The PROIEL annotation makes both testable at scale: clause-initial coordination is one query; agreement controllers and targets are another. The same queries then run on the Vulgate, Gothic, and OCS partitions, and we can ask whether the calques propagate or whether each receiving tradition restructures them.

For the Hebrew studies side of the room: this is the entry point. The Hebrew source of LXX is not in AthDGC (we are not a Hebrew corpus). But the AthDGC LXX partition is the immediate downstream witness of that Hebrew source, and the syntactic profile of the LXX is, in part, a profile of its Hebrew source as filtered through the translators’ choices. Collaboration on that filter is one of the things I would like this symposium to seed.

12c. The Septuagint, a concrete example

Genesis 1:3-4 (the fiat lux sequence)

Heb. וַיֹּ֥אמֶר אֱלֹהִ֖ים יְהִ֣י א֑וֹר וַֽיְהִי־א֖וֹר׃ וַיַּ֧רְא אֱלֹהִ֛ים אֶת־הָא֖וֹר כִּי־טֽוֹב

Tr. wa-yyōʾmer ʾĕlōhîm yəhî ʾôr wa-yhî ʾôr. wa-yyarʾ ʾĕlōhîm ʾeṯ ha-ʾôr kî ṭôḇ.

LXX καὶ εἶπεν ὁ θεός Γενηθήτω φῶς. καὶ ἐγένετο φῶς. καὶ εἶδεν ὁ θεὸς τὸ φῶς ὅτι καλόν.

Gloss and said the god let-become light. and became light. and saw the god the light that good.

Three Hebrew wayyiqtol forms (consecutive narrative tense) are calqued by three clause-initial καί + finite verb in the LXX. Classical Greek prose would subordinate (participle, μέν / δέ, ὅτε clause).

This is the slide where the audience sees the hypothesis in action on a sentence they know. Hebrew waw-consecutive (the wayyiqtol form) is the Hebrew narrative backbone: clause after clause after clause, each beginning with the conjunction waw plus a prefix-conjugation verb. Greek does not work that way. Classical Greek prose chains clauses by subordination, with participles, with μέν / δέ, with temporal clauses. The Septuagint translators choose a third option: they calque the Hebrew structure directly, with clause-initial καί + finite verb. That is a Hebrew structure wearing Greek vocabulary. And the New Testament writers, who learn their literary Greek from the Septuagint, absorb the same pattern. So this small καί propagates into Mark, into John, and from there into the Vulgate, the Wulfila Gothic, the Old Church Slavonic Marianus, and into Classical Armenian. The same καί. The same Hebrew grammatical structure, travelling across five Indo-European branches by way of the Septuagint.

12d. Three thousand years, one schema

The eight periods of Greek on the AthDGC PROIEL spine:

One picture. Three thousand years of Greek, eight periods, all on the same PROIEL XML 2.0 schema. The Koine box is highlighted because that is where the Septuagint and the New Testament sit, and that is the hinge of the whole chain: a Greek translation that absorbs a Hebrew structural calque and then exports it eastward into Wulfila, into Marianus, into the Armenian Bible.

13. What this offers the intertextuality programme

1. A test bed: 3000 years on one schema, queryable across periods.
2. Structural detection on top of surface detection: not “the same words were re-used”, but “the same propositional content was re-encoded into a different syntax”.
3. An empirical handle on the parallel-coexistent-grammars hypothesis: every retelling and retranslation is a controlled minimal pair.
4. Reusable infrastructure: the same pipeline can run on Hebrew, Sanskrit, Tibetan, with the relation inventory adjusted upstream.

The fourth point is where I would like to find collaborators in this room. The pipeline is open. The metadata register is open. The schema is the PROIEL schema, which is already used outside Greek for early Indo-European; nothing in it is Greek-specific.

14. Resources

• Public site: https://athdgc.github.io
• Corpus repository: https://github.com/AthDGC/athdgc-corpus (under v0.5 audit)
• Tools organisation: https://github.com/AthDGC
• Live PROIEL viewer: https://dialing.enl.uoa.gr/proiel/
• arXiv preprint: cs.CL (June 2026)
• Zenodo concept DOI: 10.5281/zenodo.20439182

Contact: nikolaos.lavidas@gmail.com · nlavidas@enl.uoa.gr

Διαχρονία Γλώσσας :Χρόνος · Athens Digital Glossa Chronos

Thank you. I am happy to talk about pipeline reuse for Hebrew, Sanskrit, or Tibetan corpora, about the parallel-coexistent-grammars hypothesis, or about the technical side of running PROIEL annotation at this scale.

Funding

Funded by the Hellenic Foundation for Research and Innovation (HFRI) under the 3rd Call for HFRI Research Projects to support Post-Doctoral Researchers, Project No. 20577; with complementary support from the Greece 2.0 National Recovery and Resilience Plan.

Compute supplied by GRNET ARIS (the Greek national high-performance computing cluster), allocation pa260305.

Corpus licence CC BY 4.0. Tooling licences MIT / Apache 2.0 / BSD.