LLM-RAG — Endangered and extinct languages (Kagoshima 2026)

LLM-RAG

Endangered and extinct languages

Setup — Kagoshima Linguistics Workshop 2026

Kagoshima Linguistics Workshop 2026

Translation and Coptic language revival

So Miyagawa 宮川創

University of Tsukuba 8 November 2026

Kagoshima University of Tsukuba

Click Next to continue.

⏱ 00:00–00:59 ｜冒頭はゆっくり。発表全体の二つの問いを示す。

Thank you very much for the invitation. Today I would like to discuss what language models can do when we give them access to linguistic resources. I will show translation experiments involving Yoron, Truku, Middle Egyptian and Coptic. The results lead to a second question: how can these tools support people who want to learn a language and write new texts? For Coptic, that question has a long history. I will discuss its declining use as an everyday language, its continued liturgical use, and modern revival activities. I will then introduce examples from our Göttingen presentation with Mario Beshara and demonstrate SINUHE.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit
Cover image: https://library.aucegypt.edu/rare-books-special-collections/collections/coptic-studies
Design image: The Metropolitan Museum of Art, Coptic Textile Fragment, late 3rd–5th century, 89.18.95. Public domain. https://www.metmuseum.org/art/collection/search/443300

Languages and the evidence available

FileViewGoFavoritesProgramsHelpYoron Island Xiulin (Truku)Egypt Meroë

Address

C:\Kagoshima\Languages

Language	Evidence used here
Yoron / Truku	Parallel texts and dictionaries
Egyptian / Coptic	Editions, lexicons and annotated texts
Meroitic	Inscriptions and editorial restorations

Evaluation

Resources differ. Evaluation must fit the task.

Miyagawa’s studies (2025–2026)

Introduction

02 / 40

⏱ 00:59–01:58 · Clicks: 4 ｜表の各行を指す。言語間の単純な優劣比較はしない。

The resources available to these projects differ considerably. In Yoron and Truku, we can work with parallel texts and dictionaries, and speakers can evaluate what a translation actually means. For Egyptian and Coptic, we also have editions, lexicons and annotated texts. Coptic has continuing liturgical users and people producing new writing. Meroitic presents a further difficulty because many aspects of its interpretation remain unsettled. I therefore use different evaluation questions for different tasks. A translation score, a plausible restoration and a useful learning experience each tell us something different. Egyptian and Coptic are also historically related, which matters when we interpret the comparison.

Map (optional): Yoron Island · Xiulin Township · Egypt · Meroë — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://escholarship.org/uc/item/22r6s881

Fluent output, wrong meaning

FileViewGoFavoritesProgramsHelpYoron Island

Yoron	Reported output	Meaning
ピジュル	daytime / beer	cold
フギチャニ	chicken	sugarcane seedlings

Lexical check

Fluency can hide a lexical error.

Miyagawa (2026a)

Introduction

03 / 40

⏱ 01:58–02:56 · Clicks: 3 ｜左から右へ二例を指す。

Let me begin with two small examples from the Yoron experiment. The expression pijuru means cold. Some systems produced words corresponding to daytime or beer. In the second row, fugichani refers to sugarcane seedlings, but one output gave chicken. These are quite specific lexical mistakes. The surrounding sentence may still sound natural, so fluency alone makes them difficult to detect. These examples come from particular recorded outputs. They do not establish a permanent ranking of model families. What they help us ask is whether access to an appropriate dictionary entry or parallel example can correct the meaning in this particular passage.

Map (optional): Yoron Island — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

RAG: a question, sources and an answer

FileViewGoFavoritesProgramsHelp

Retrieval-augmented generation

1Question

2Retrieve

3Generate

4Answer

Dictionary
Parallel texts

Question +
selected passages

Text for the
reader to check

Lewis et al. (2020); explanatory schematic

RAG

04 / 40

⏱ 02:56–03:52 · Clicks: 4 ｜左から順に指す。検索対象の資料、選ばれた箇所、生成された答えを区別する。RAGは検索拡張生成と一度だけ補足。

RAG stands for retrieval-augmented generation. Think of giving a model a small reading packet before asking it to answer. A question first guides a search through the selected collection. The search returns relevant passages, such as a dictionary entry or a parallel sentence. The model then receives both the question and those passages, and writes an answer. The important addition is access to evidence at the moment of answering. The collection can contain materials that the model did not reliably learn during training. In the experiments discussed here, we supply retrieved text without updating the model weights.

Sources:
Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. https://arxiv.org/abs/2005.11401

A worked example: pijuru

FileViewGoFavoritesProgramsHelpYoron Island

Question:

Yoron word

ピジュル

Retrieved clue

pijuru: cold

Answer

Here, pijuru
means “cold”.

The sentence determines which sense fits.

Illustration based on the Yoron example. The retrieval packet is schematic.

RAG

05 / 40

⏱ 03:52–04:50 · Clicks: 3 ｜ピジュルを発音し、中央の資料から右の答えへ指す。検索ログの再現ではなく、既存の与論語例で仕組みを説明する図。

Here is a schematic example using the Yoron word pijuru from the earlier lexical example. We ask what the word means. Retrieval brings an entry or a relevant parallel example into the reading packet. For teaching purposes, the middle of this slide shows the useful gloss, cold. The model can use that clue to produce an answer such as, here pijuru means cold. This is an illustration of the mechanism, not a reproduction of the actual prompt or retrieval log. In a translation task, we retrieve evidence for a sentence and still have to choose the sense that fits its context.

Map (optional): Yoron Island — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
Illustrative walkthrough based on the Yoron lexical example in Miyagawa (2026a). The packet and output wording are schematic.
Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. https://arxiv.org/abs/2005.11401

Three checks on a RAG answer

FileViewGoFavoritesProgramsHelp

Question	Evidence to inspect
Right source?	The original entry or passage
Faithful answer?	The answer beside its source
Enough evidence?	What the source leaves unresolved

Reading

The reader checks the answer against the source.

Lewis et al. (2020); explanatory schematic

RAG

06 / 40

⏱ 04:50–05:49 · Clicks: 4 ｜検索の誤り、生成の誤り、資料の不足を一つずつ説明する。資料を表示すれば自動的に正解になるわけではないことを短く述べる。

Access to a source helps us inspect an answer, but it does not settle the interpretation. First, did the search find the right entry or passage? A similar spelling can lead to the wrong word. Second, does the answer follow the evidence it received? A model may still ignore a relevant passage or add unsupported information. Third, is the source sufficient for this question? A dictionary gloss alone may not resolve a grammatical or historical problem. We therefore keep the source available to the reader and record what remains uncertain. These checks connect the diagram to the lexical examples and evaluation results that follow.

Sources:
Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. https://arxiv.org/abs/2005.11401
Checks on retrieval, evidence use and source sufficiency are the presenter’s evaluation framework.

Yoron: aligned phrases

FileViewGoFavoritesProgramsHelpYoron Island Amami Islands

Retrieval collection

9,007

aligned phrase pairs

38 tales584 sentences

Yoron	Japanese
シマ・ヌ	島・の
パジマイ	はじまり

Test set

Test: 14 sentences from another publication.

Kiku (1985, 2017), Miyagawa (2026a)

Yoron · Truku · Egyptian

07 / 40

⏱ 05:49–06:43 · Clicks: 3 ｜ 9,007と右の対訳例を指す。

For Yoron, the retrieval collection draws on Kiku Chiyo’s folk tales with Japanese translations. We organized 38 tales, containing 584 sentences, into 9,007 aligned bunsetsu pairs. The table gives a very small example of that alignment. The phrase boundaries let the system retrieve a compact unit with a corresponding Japanese expression. In this configuration, it retrieves six similar pairs for a query. The evaluation uses fourteen sentences from another publication by Kiku Hidefumi. Separating the publications helps us avoid testing only on the same material, although fourteen sentences remain a very small sample.

Map (optional): Yoron Island · Amami Islands — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Yoron: translation results

FileViewGoFavoritesProgramsHelpYoron Island

BLEU

35.7

No retrieval

45.6

With retrieval

+9.9

BLEU points

14 sentences

Miyagawa (2026a), BLEU 0–100

Yoron · Truku · Egyptian

08 / 40

⏱ 06:43–07:42 · Clicks: 4 ｜棒を左から右へ指す。

Here is the comparison within Gemini 3.0 Pro Preview. Without the retrieval collection, the reported BLEU score was 35.7. With retrieval, it was 45.6, a gain of 9.9 points. Other reported measures also improved. This is encouraging because the model identity stays constant in the comparison. The size of the test remains an important limitation, however. We need more sentences and more genres before estimating general translation performance. I would also like to inspect which retrieved pairs contributed to each correction. The aggregate score is useful as an overview, but the individual lexical and grammatical decisions are often more informative for a linguist.

Map (optional): Yoron Island — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Yoron: what improves?

FileViewGoFavoritesProgramsHelpYoron Island

Expression	With retrieval
ピジュル	cold
フギチャニ	sugarcane seedlings

Grammatical information

EvidentialityKeep the distinction in a gloss.

Miyagawa (2026a), Kibe (2019)

Yoron · Truku · Egyptian

09 / 40

⏱ 07:42–08:40 · Clicks: 3 ｜語彙の改善から証拠性へ話を移す。

The lexical examples show what some of that improvement looks like. With retrieval, the output gives cold for pijuru and sugarcane seedlings for fugichani. Grammatical information requires a different kind of assessment. For example, an evidential distinction in Yoron may have no equally obligatory grammatical expression in Japanese. A natural translation can leave it implicit. If the aim is linguistic analysis or teaching, an explanatory gloss can make the distinction visible. I would therefore distinguish a readable translation from an analysis that preserves the relevant grammatical information. We should evaluate both when a tool is intended for learners or researchers.

Map (optional): Yoron Island — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Mandarin into Truku

FileViewGoFavoritesProgramsHelpXiulin (Truku)Wanrong Zhuoxi Taroko Gorge

BLEU

4.35

No retrieval

7.47

With retrieval

Bible pairs
+ dictionary

6 paragraphs
1 reference

Evaluation

Speaker review remains essential.

Miyagawa, Tang, Sato & Koizumi (2026)

Yoron · Truku · Egyptian

10 / 40

⏱ 08:40–09:38 · Clicks: 6 ｜スコアと小さなテスト規模の両方を示す。

The Truku experiment concerns translation in the other direction: from Mandarin into Truku. The collection combines aligned Bible verses with dictionary entries. In this pilot, BLEU increases from 4.35 to 7.47. The test consists of a textbook preface with six paragraphs and one reference translation, so the numbers have a narrow scope. The remaining errors require careful work with speakers, particularly in voice and argument structure. A Bible-based collection can provide valuable examples while also favoring a particular register. This project involves Apay Ai-yu Tang, Manami Sato, Masatoshi Koizumi and Truku collaborators. Their linguistic and community knowledge is central to evaluation.

▶ Click 5 → pop-up window: Truku Translator AI (Live app on Google AI Studio)
▶ Click 6 → YouTube pop-up: 原視新聞網 TITV News — “族語學習新亮點 AI建構太魯閣族語言模型” (New highlight in Indigenous-language learning: AI builds a Truku language model)
The next click closes it and continues. To reopen: ← then →, or the Favorites menu (“Show here”). Videos play inside the slide only when the deck is opened from a web address (e.g. http://localhost:8000/…); opened as a local file, the pop-up offers “Play on YouTube”, which opens the video in its own window.

Map (optional): Xiulin Township · Wanrong Township · Zhuoxi Township · Taroko National Park — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Middle Egyptian: tpj šmꜥw

FileViewGoFavoritesProgramsHelpElephantine

tpj šmꜥw

Without retrieval	With retrieval
upon Upper Egypt	first nome of Upper Egypt

Score

Thutmose I decree: 32.46 to 35.43 SacreBLEU.

Miyagawa (2025), NLP4DH

Yoron · Truku · Egyptian

11 / 40

⏱ 09:38–10:36 · Clicks: 3 ｜ tpj šmꜥwと訳語の差を指す。

The Middle Egyptian example comes from a decree of Thutmose the First. The expression tpj šmꜥw is the important part here. In the comparison output, the phrase became a general reference to being upon Upper Egypt. The retrieval-assisted output instead follows the reference interpretation, the first nome of Upper Egypt, in connection with Elephantine. Across this decree, SacreBLEU rises from 32.46 to 35.43. I find the local interpretation more revealing than the size of that increase. It shows how a lexical or administrative distinction can affect the reading of a historical text. The evaluation concerns one decree and one reference translation.

Map (optional): Elephantine, Aswan — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Coptic: changing domains of use

FileViewGoFavoritesProgramsHelpEgypt

GeneralDomainsVitalityHistory

Coptic

Language shift
and continuity

Domain	Trajectory
Everyday speech	Shift to Arabic
Writing	Arabic expands
Liturgy	Coptic continues
Learning	Teaching and new writing

Richter (2023), Teepe (2025)

Coptic: history & revival

12 / 40

⏱ 10:36–11:30 · Clicks: 4 ｜日常語と典礼語を別々の行として示す。

Before turning to Coptic language technology, I want to distinguish the domains shown here. Everyday speech shifted toward Arabic, while Coptic remained in worship and learned practice. Writing has its own chronology. Teaching can sustain competence after household transmission has contracted, and new writing can create additional uses. These distinctions change what we mean when we describe Coptic as extinct or revived. They also change the appropriate goal of a digital tool. I will use revival here for activities intended to widen use, while keeping the scale of everyday transmission an open empirical question.

Map (optional): Egypt — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://escholarship.org/uc/item/22r6s881
https://elischolar.library.yale.edu/gsas_dissertations/1663/
https://doi.org/10.5356/jorient.58.2_184

Coptic: endangered, with continuing use

FileViewGoFavoritesProgramsHelpEgypt

GeneralDomainsVitalityHistory

Ethnologue 2026*

Endangered

Transmission	Continuing uses
No L1 speakers reported*	Liturgy, learning and new writing

*Reported by Purushothama et al. (2026), citing Ethnologue.

Ethnologue (2026), cited in Purushothama et al. (2026), n. 2

Coptic: history & revival

13 / 40

⏱ 11:30–12:32 · Clicks: 1 ｜ endangeredを「認定」ではなく分類として説明。Ethnologueの記載は2026年論文による間接確認。母語伝承・典礼・再学習を分ける。

Ethnologue’s 2026 classification is cited in a recent ACL paper as endangered, with no first-language speakers. I am citing the published paper here: the live Ethnologue entry was not accessible for direct verification. This label concerns language vitality. It does not mean that Coptic has ceased to be used. Liturgical recitation, language teaching, conversation among learners and new writing have different histories and different measures of continuity. We should therefore ask who uses Coptic, for which activities, and how the ability to use it is passed on. Claims about individual families raising children in Coptic also require careful documentation; they cannot be converted into a reliable population total.

Map (optional): Egypt — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
Purushothama et al. (2026), p. 36177, n. 2; citing Eberhard, Simons & Robinson, Ethnologue, 29th edition (2026).
https://aclanthology.org/2026.findings-acl.1803.pdf
https://www.ethnologue.com/language/cop/

The shift to Arabic

FileViewGoFavoritesProgramsHelpEgypt

GeneralDomainsVitalityHistory

Period	Change
7th century onward	Gradual language shift
9th–11th centuries	Arabic use expands
12th–14th centuries	Coptic loses written domains

Last speaker?

The “last speaker” date remains uncertain.

Richter (2009, 2023)

Coptic: history & revival

14 / 40

⏱ 12:32–13:30 · Clicks: 4 ｜年代の幅を示す。単一の消滅年は述べない。

The shift to Arabic unfolded over centuries. The seventh-century conquest provides a political starting point, but it does not give us a single date for language loss. Arabic use expanded unevenly, with a stronger shift in later medieval centuries. Christian Arabic writing and translation grew as Coptic lost many written domains. Surviving texts do not tell us directly what each family spoke. Reports of much later Coptic speakers also need historical scrutiny. For that reason, I do not assign a precise last-speaker year. The table separates broad stages of change, rather than presenting a sudden replacement of one language by another.

Map (optional): Egypt — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://doi.org/10.1017/CBO9780511641992.019
https://escholarship.org/uc/item/22r6s881

Learning Coptic through Arabic

FileViewGoFavoritesProgramsHelpCairo

Muqaddimāt

Coptic grammar explained in Arabic

1240s–1260s

6grammars

100+manuscript copies

Teepe (2025)

Coptic: history & revival

15 / 40

⏱ 13:30–14:27 · Clicks: 3 ｜ 6文法書と100以上の写本を指す。

One particularly interesting response was the production of Coptic grammars in Arabic, called muqaddimāt. Ramona Teepe’s research identifies six grammars composed between the 1240s and 1260s, with more than a hundred surviving manuscript copies. Their authors addressed Arabic-reading learners and treated Coptic as a language that could be acquired through explicit instruction. This gives us a concrete example of adaptation in transmission. My connection with the present project is pedagogical: learners need explanations that make forms and patterns accessible. The medieval evidence also reminds us that Arabic participation in Coptic learning was part of the language’s continuing intellectual life.

Map (optional): Cairo — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://elischolar.library.yale.edu/gsas_dissertations/1663/

Teaching and liturgical reading

FileViewGoFavoritesProgramsHelpCairo AUC Library

AUC Coptic Studies

collection image (illustration)

ʿIryān Muftāḥ

1826–1886

Clerical College

Grammar manuals

Greco-Bohairic reading

AUC Libraries, Coptic Studies collection

Coptic: history & revival

16 / 40

⏱ 14:27–15:23 · Clicks: 3 ｜画像はAUCコレクションの参考図。特定の文法書とは紹介しない。

In the nineteenth century, education and pronunciation reform became closely connected. Iryan Moftah taught at the Clerical College established under Cyril the Fourth. His grammar manuals and liturgical publications supported teaching and the Greco-Bohairic reading tradition. The image illustrates the AUC Coptic Studies collection. I am not identifying this particular manuscript page as one of his grammar manuals. For our purposes, the distinction between a pronunciation tradition and a grammatical variety matters. When software offers a spoken form, it should tell learners which reading convention it follows. That question is separate from the reconstruction of ancient pronunciation.

Map (optional): Cairo · American University in Cairo — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://library.aucegypt.edu/rare-books-special-collections/collections/coptic-studies
https://escholarship.org/uc/item/22r6s881

Labīb and spoken Coptic

FileViewGoFavoritesProgramsHelpCairo

Iqlawdiyūs Labīb

Project	Aim
ʿAyn Shams	Spoken Coptic
Egyptian heritage	Cultural identity

Labīb · portrait

from his lexicon · public domain

Miyokawa (2016). Portrait: Labīb lexicon, public domain

Coptic: history & revival

17 / 40

⏱ 15:23–16:19 · Clicks: 2 ｜ Labībの肖像を示し、雑誌名を発音する。

Here is Iqlawdiyūs Labīb, also known as Claudius Labib. Around the turn of the twentieth century, he advocated renewed spoken use of Coptic. Hiroko Miyokawa examines this effort through his magazine Ayn Shams. Labib connected Coptic with ancient Egyptian heritage and with claims about Egyptian identity. His campaign did not establish a broad vernacular community, but its cultural argument mattered. I include it because a revival project can have several aims: a wider readership, a place in education, or a claim about heritage. We should ask which aim participants are pursuing before deciding how to evaluate an application.

Map (optional): Cairo — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://doi.org/10.5356/jorient.58.2_184
https://commons.wikimedia.org/wiki/File:Claudius-labib-photo.jpg

Coptic inside church classrooms

FileViewGoFavoritesProgramsHelpMinya Nazlet al-Fallahin Taha al-Amidah Samalut

Church setting	Since
Virgin Mary Church Nazlet al-Fallahin, Minya	2007
Taha al-Amidah Samalout diocese	2024

Documented practice

Reading

Writing

Conversation

Full-immersion curriculum
(not documented)

Coptic-language lectures at an annual forum

Communicative use is documented; a full-immersion curriculum is not.

Milad (2025), Al Manassa: interviews with the teacher and learners

Coptic: history & revival

18 / 40

⏱ 16:19–17:21 · Clicks: 5 ｜ Polisは扱わない。正教会主体の2教室と、実際に確認できる会話・コプト語講演を紹介。全授業コプト語のイマージョン課程とは断定しない。

A concrete example comes from the Virgin Mary Church at Nazlet al-Fallahin in Minya. Gergis Ibrahim established its Coptic school in 2007. In 2024 he helped establish another in Taha al-Amidah, in the Samalout diocese, with Father Lukas Nabil. An interview report describes pupils reading, writing and holding simple conversations in Coptic. It also reports Coptic-language lectures at an annual forum, and former students becoming teachers. These are documented opportunities to learn through using the language. The report does not establish that every lesson, or the whole curriculum, is conducted exclusively in Coptic. I therefore describe these as church-led communicative practices, without claiming a verified full-immersion programme.

Map (optional): Minya · Nazlet al-Fallahin · Taha al-Amidah · Samalut — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
Antoun Milad (2025), The Coptic language: Breathing life into the voice of ancient Egyptians, Al Manassa, 19 July. Firsthand interviews.
https://manassa.news/en/stories/25470
https://coptic-treasures.com/place/schools/coptic-language-school-menya/

Revival as an activity

FileViewGoFavoritesProgramsHelpGöttingen

Address

Prototype screen — Göttingen presentation

Beshara & Miyagawa (2026), prototype

Coptic: history & revival

19 / 40

⏱ 17:21–18:17 ｜試作画面であることを一言添える。

This brings us to actual activities. The image shows a learning-dashboard prototype included in the Göttingen presentation. It is useful here because it places language material in a setting where somebody reads and practices. It does not provide evidence of measured learning gains. For a revival project, I would want to know whether people return to the material, understand it, and begin using patterns in their own writing. Translation scores cannot answer those questions on their own. We need observations of learners and readers over time, with goals that reflect what the participants themselves want to do.

Map (optional): Göttingen — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://library.aucegypt.edu/rare-books-special-collections/collections/coptic-studies
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Göttingen with Mario Beshara

FileViewGoFavoritesProgramsHelpGöttingen

Göttingen

13th International Congress of Coptic Studies

31 July 2026

THOTH AI 2.0

Translation and analysis

Coptic Composer AI

Composition and revision

Beshara & Miyagawa (2026)

Göttingen: THOTH AI & Composer

20 / 40

⏱ 18:17–19:13 · Clicks: 3 ｜共同発表の場と二人の役割を簡潔に紹介する。

Mario Beshara and I presented this work at the Thirteenth International Congress of Coptic Studies in Göttingen on the thirty-first of July, 2026. Our presentation brought together THOTH AI 2.0 and a Coptic composition interface. The two sides of the table summarize the distinction. Translation and analysis can help us read existing texts. Composition and revision concern producing something new in Coptic. Mario’s practical work is especially valuable for the second task because it gives us texts to examine and an author’s decisions to discuss. I will show one example of that process after the translation comparison.

▶ Click 3 → pop-up window: THOTH AI (Live app · may ask for the sign-in first)
The next click closes it and continues. To reopen: ← then →, or the Favorites menu (“Show here”). Videos play inside the slide only when the deck is opened from a web address (e.g. http://localhost:8000/…); opened as a local file, the pop-up offers “Play on YouTube”, which opens the video in its own window.

Map (optional): Göttingen — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit
https://trios.tsukuba.ac.jp/en/researcher/0000004893

Bohairic: Life of Shenoute §88

FileViewGoFavoritesProgramsHelpWhite Monastery Wadi el-Natrun

Bohairic into English · BLEU

31.4

Gemini interface

40.63

THOTH AI 2.0

1 passage
1 reference

Different
interfaces

Miyagawa (2026c), BLEU 0–100

Göttingen: THOTH AI & Composer

21 / 40

⏱ 19:13–20:08 · Clicks: 3 ｜ Bohairicの評価範囲を短く説明する。

This comparison uses one passage from the Bohairic Life of Shenoute, section eighty-eight, translated into English. The reported BLEU scores are 31.40 for the Gemini interface and 40.63 for THOTH AI 2.0. There is one reference translation. Although the underlying model label overlaps, the deployed interfaces and prompts differ, so this comparison does not isolate retrieval as the only cause of the difference. It records what the evaluated systems produced on this passage. For a controlled ablation, we would need to hold the prompt and other settings constant, archive the retrieved passages, and compare repeated runs.

Map (optional): White Monastery · Wadi el-Natrun — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Coptic: lexical and factual errors

FileViewGoFavoritesProgramsHelp

ϭⲁⲗⲟϫ

Reference / THOTH	Other outputs
big toe	heel

Content check

Invented events need a separate check.

Miyagawa (2026c)

Göttingen: THOTH AI & Composer

22 / 40

⏱ 20:08–21:02 · Clicks: 3 ｜コプト語を指してからbig toeとheelを比較する。

A specific word makes the comparison easier to assess. The Coptic form shown here corresponds to big toe in Bell’s reference translation, and THOTH AI gave the same expression. Several other outputs used heel. That is a local lexical decision we can check. A different output added an event in which Christ appeared in a dream, although the passage contains no such event. I distinguish these error types because their consequences differ. Lexical variation, mistranslation and invented content should receive separate attention in a philological evaluation. A single translation score tends to combine them.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Beshara’s translation of a Grimm tale

FileViewGoFavoritesProgramsHelp

Grimm tale · Bohairic

Beshara & Miyagawa (2026), slide 6

THOTH AI draft

Beshara’s correction
and editing

New Coptic reading material

Beshara & Miyagawa (2026), slide 6

Göttingen: THOTH AI & Composer

23 / 40

⏱ 21:02–21:59 · Clicks: 4 ｜ Marioさんの編集・校正の役割を強調する。

This page comes directly from Mario Beshara’s translation of a Grimm tale, as reproduced in our Göttingen slides. THOTH AI assisted the draft, and Mario corrected and edited the Coptic. The result is new reading material beyond the inherited liturgical and literary corpus. What interests me is the author’s revision process. Which suggestions does he retain? Which grammatical patterns need correction? Can readers follow the story without continual explanation? Those questions would let us evaluate the support offered by the tool in a concrete activity. The example documents a translation practice. Wider claims about revitalization would require further evidence.

▶ Click 4 → YouTube pop-up: Coptic Tongue — “Ϯϣⲟⲙⲧ ⲛ̀Ⲁⲥⲡⲓ” (‘The Three Languages’)
The next click closes it and continues. To reopen: ← then →, or the Favorites menu (“Show here”). Videos play inside the slide only when the deck is opened from a web address (e.g. http://localhost:8000/…); opened as a local file, the pop-up offers “Play on YouTube”, which opens the video in its own window.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Coptic Composer AI

FileViewGoFavoritesProgramsHelp

Lexicon search

Text editor

AI assistant

Lexicon lookup alongside the editor

Beshara & Miyagawa (2026), prototype

Göttingen: THOTH AI & Composer

24 / 40

⏱ 21:59–23:24 · Clicks: 3 ｜左の辞書、中央の編集欄、右の操作欄を順に指す。

The Göttingen presentation also showed this Coptic Composer interface as a prototype under construction. The lexicon search sits beside the editor, with a virtual keyboard below. Assistant controls appear on the right. The arrangement makes the dictionary available while the author is writing, which is a useful design choice for a learner or translator. I am showing the development stage presented at the congress. The visible controls should not be read as proof that every function is currently deployed. SINUHE, which I will show next, is a separate application whose public interface I inspected directly.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

SINUHE: four writing modes

FileViewGoFavoritesProgramsHelp

Mode	Input	Support
Hieroglyphic	Readings / Gardiner codes	Layout and dictionaries
Hieratic	Unicode + hieratic font	Historical form links
Demotic	Readings / English meanings	Spelling variants
Coptic	Roman keys / dialects	Glosses and suggestions

Hieratic and Demotic: very experimental

Work in progress

THOTH AI is currently being integrated into SINUHE.

HIERO.AVI - Media Player

0:00 / 0:00

Screen recording

SINUHE public interface, observed 7 November 2026

SINUHE

25 / 40

⏱ 23:24–24:46 · Clicks: 5 ｜ 4行を順に指す。ヒエラティックとデモティックは実験段階と一言添える。

The media player on the right follows the rows: each row switches to a screen recording of that mode (recorded 8 Nov 2026). Last step: THOTH AI integration note.

SINUHE brings four writing modes into one workspace. Hieroglyphic mode combines readings and Gardiner codes with Unicode input, layout controls and dictionary lookup. Hieratic mode uses the same encoded text with a hieratic display font and links to historical sign forms. Demotic mode adds word and spelling selection, including searches from English meanings. Coptic mode provides dialect choices, diacritics, lexical analysis and corpus-based continuations. These modes serve different writing tasks. The application labels Hieratic and Demotic as very experimental, so I distinguish their current functions from a complete account of historical handwriting. [added] We are also currently integrating THOTH AI into SINUHE; this work is still in progress.

Sources:
https://sinuhe.vercel.app/

SINUHE: hieroglyphic input and layout

FileViewGoFavoritesProgramsHelp

Input	Compose	Consult
Reading or code	Groups and direction	Lexicon and corpus

Cartouches, editable files and PNG export

↓ live demo

SINUHE public interface, observed 7 November 2026

SINUHE

26 / 40

HIERO.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 24:46–26:15 · Clicks: 5 ｜ nfrの入力画面を指し、下の3列へ移る。実演ではDirectionとCartoucheのメニューを開く。

Last step opens a recorded demo (HIERO.AVI: nfr → insert → next word → Direction → Cartouche, ~24 s, muted, loops). Talk over it, then press → for the next slide.

Here I type nfr and convert it to a hieroglyphic sign. A Gardiner code can also be used. Vertical and horizontal joiners, together with overlay controls, let the user compose sign groups. The Direction menu offers horizontal writing in both directions, vertical columns, and choices for the facing of signs. Selected text can be placed inside a cartouche or another enclosure. Dictionary glosses and corpus suggestions remain available alongside the editor. An editable file preserves work for later revision, while PNG export preserves its displayed appearance. These are text-entry and editing functions. The user still selects and checks the intended reading and spelling.

Sources:
https://sinuhe.vercel.app/

SINUHE: hieratic display and references

FileViewGoFavoritesProgramsHelp

nfr

EZGlyph hieratic

Unicode text
Font-based display

Function	Use
Input	nfr / Dd / G17
References	AKU-PAL Thot / Möller
Fallback	Hieroglyphic forms

Experimental. Check historical forms against sources.

SINUHE public interface, observed 7 November 2026

SINUHE

27 / 40

HIERATIC.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 26:15–27:12 · Clicks: 5 ｜画面から切り出したnfrの字形を示す。フォント表示と、時代・資料ごとの字形資料を区別する。

Last step opens a recorded demo (HIERATIC.AVI: nfr in hieratic mode, ~6 s, muted, loops). Talk over it, then press → for the next slide.

In Hieratic mode, the same nfr input appears in the built-in EZGlyph hieratic font. The underlying text remains Unicode. This makes it possible to change the display while keeping an encoded transcription. The application also provides reference panels for AKU-PAL, Thot and Möller. In the AKU-PAL panel, users can find a sign, follow a link to its historical forms, and insert a supported Unicode transcription. A font rendering should not be treated as a facsimile of a particular manuscript. Where the font has no hieratic form, the application falls back to a hieroglyphic form. The current mode is very experimental.

Sources:
https://sinuhe.vercel.app/
https://aku-pal.uni-mainz.de/
https://ezglyph.com/ezglyph_005.htm

SINUHE: Demotic words and spellings

FileViewGoFavoritesProgramsHelp

English

“”

Demotic

Compare spellingsCDD / DWL / Erichsen / TLA

Experimental. Candidates link to source evidence.

SINUHE public interface, observed 7 November 2026

SINUHE

28 / 40

DEMOTIC.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 27:12–28:42 · Clicks: 5 ｜ waterからmwへの候補列を示す。頻度はこの検索で表示された綴りの頻度。CDDは原本参照機能であり、OCR機能とは説明しない。

Last step opens a recorded demo (DEMOTIC.AVI: English → Demotic, “water”, ~9 s, muted, loops). Talk over it, then press → for the next slide.

Demotic mode supports searches by transliteration and by English meaning. In this example, I typed water and opened English to Demotic. The interface offered mw and several recorded spellings, with corpus frequencies and links to DPDP palaeographic evidence. The user can compare the candidates before inserting one. Other panels provide access to CDD, DWL, Erichsen and TLA forms. The CDD panel locates and opens original PDF pages. It is a page-reference library, and does not turn printed Demotic signs into editable text. The displayed forms use the dpdp font distributed by ELRC. The mode remains very experimental, and readings and spellings require source checking.

Sources:
https://sinuhe.vercel.app/
https://elrc.dev/fonts/dpdp.woff

SINUHE: next-word suggestions

FileViewGoFavoritesProgramsHelp

nfr

Next-word corpus

Earlier Egyptian · TLA v18

Next word	Corpus count
wr.t	43 / 225
n	17 / 225
ḫpi̯	16 / 225

SINUHE, observed 7 November 2026

SINUHE

29 / 40

HIERO.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 28:42–29:36 · Clicks: 4 ｜ 43 / 225を指す。コーパス頻度と生成モデルの確率を区別する。

Last step opens a recorded demo (HIERO.AVI segment: nfr → suggestions → wr.t, ~8 s, muted, loops). Talk over it, then press → for the next slide.

Let us look more closely at the suggestions after nfr. With the Earlier Egyptian TLA version eighteen corpus selected, the interface showed wr.t forty-three times among 225 recorded continuations. It showed n seventeen times and ḫpi̯ sixteen times. These numbers describe corpus occurrences displayed by the application. They are not probabilities generated by a language model. A writer can inspect the examples and decide whether a continuation fits the new sentence. Frequent combinations may reflect formulae or particular genres. Corpus selection therefore matters, and a common continuation still requires a grammatical and contextual decision.

Sources:
https://sinuhe.vercel.app/
https://huggingface.co/datasets/thesaurus-linguae-aegyptiae/tla-Earlier_Egyptian_original-v18-premium

SINUHE: Coptic writing support

FileViewGoFavoritesProgramsHelp

“his name”

Next word	Count
ⲡⲉ	5 / 27
ⲛⲁⲱⲛϩ	2 / 27

Coptic SCRIPTORIUM

↓ live demo

SINUHE and Coptic SCRIPTORIUM, observed 7 November 2026

SINUHE

30 / 40

COPTIC.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 29:36–30:34 · Clicks: 5 ｜語構造、候補、出典リンクを順に指す。実演は30秒程度。

Last step opens a recorded demo (COPTIC.AVI: pefran → ⲡⲉϥⲣⲁⲛ → ⲡⲉ, ~11 s, muted, loops). Talk over it, then press → for the next slide.

In Coptic mode, I entered pefran, meaning his name. The interface separates the possessive element pef from the noun ran and gives grammatical labels. It then offers continuations from Coptic SCRIPTORIUM. In this observation, pe occurred five times among twenty-seven continuations, and naōnh occurred twice. Choosing a suggestion inserts it and updates the next set of candidates. The interface also offers dialect filters, spelling-related controls and dictionary links. During a demonstration I would point first to the form analysis, then to the suggestions, and finally to the source link. Those are three different pieces of information that a learner can check.

Sources:
https://sinuhe.vercel.app/

Two Coptic Orthodox churches in Japan

FileViewGoFavoritesProgramsHelpKizugawa church Ome church St Andrew’s, Tokyo

Kizugawa, Kyoto

St Mary & St Mark

Kyoto church building

Ome, Tokyo

St Mary & St John

Photo: Tokyo wedding at St Andrew’s
Published 2020 · historical venue

Official church website; Yano (2020), AMOR. Current names checked 7 November 2026

Coptic in Japan

31 / 40

⏱ 30:34–31:34 · Clicks: 2 ｜京都は現在の教会堂、東京は2020年掲載の会衆写真。東京の写真を青梅の会堂と説明しない。現行名称と旧資料の聖十二使徒教会を区別。

The church’s current public notice lists two locations. St Mary and St Mark is in Kizugawa, Kyoto, at 2-1 Saganaka Shimizu. The Tokyo notice names St Mary and St John at 1-chome-47 Katsunuma in Ome. The photograph on the left shows the Kyoto church. The image on the right records a Coptic wedding held in St Andrew’s Anglican Church in Tokyo, published in 2020. It documents the congregation’s earlier use of borrowed church space; it is not a photograph of the present Ome building. The same 2020 account called the Tokyo parish the Twelve Disciples. I keep that historical name separate from the current notice.

Map (optional): St Mary & St Mark Coptic Orthodox Church · St Mary & St John Coptic Orthodox Church · St Andrew’s Church, Shiba — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
Current locations: official church notice, accessed 7 November 2026.
https://www.japancopticchurch.com/general-5
Kyoto photograph: official church website.
https://www.japancopticchurch.com/
Tokyo historical photograph and account: Thomas Tatsuhiro Yano (2020), AMOR, 19 January; photo credited there to the church Facebook page.
https://webmagazin-amor.jp/2020/01/19/tokushu39_6/

From visiting ministry to a local community

FileViewGoFavoritesProgramsHelpKizugawa church Ome church

Year	Milestone
2004	Thomas Kanazaki begins ministry in Japan
2016	Kyoto church opens on 18 July
2017	Pope Tawadros II visits the Kyoto church
Today	Kyoto and Tokyo locations listed

Coptic · Arabic · English · Japanese

Miyokawa (2016); Kirishin (2016); Yano (2020); official church website

Coptic in Japan

32 / 40

⏱ 31:34–32:33 · Clicks: 5 ｜ 2004年の活動開始と2016年の会堂開堂を混同しない。4言語の典礼から、次の読書室開発へつなぐ。

The history begins before the Kyoto building. Father Thomas Kanazaki worked in Japan from 2004, initially based in Kobe and later serving communities in several cities. Hiroko Miyokawa describes attending liturgy in his Kobe home in 2005 and later in borrowed church premises. On 18 July 2016, the Kyoto community opened its church in a former Protestant building. Pope Tawadros the Second visited and blessed the church in 2017. Today, the public notice lists Kyoto and Tokyo. The Kyoto website describes worship using Coptic, Arabic, English and Japanese. That multilingual setting is the immediate reason for building reading support that includes Japanese.

Map (optional): St Mary & St Mark Coptic Orthodox Church · St Mary & St John Coptic Orthodox Church — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
Hiroko Miyokawa (2016), Christian Today, 13 September.
https://www.christiantoday.co.jp/articles/22032/20160913/coptic-orthodox-church-miyokawa-hiroko-1.htm
Kirishin (2016), report of the 18 July opening.
https://www.kirishin.com/2016/08/06/18204/
Thomas Tatsuhiro Yano (2020), AMOR.
https://webmagazin-amor.jp/2020/01/19/tokushu39_6/
https://www.japancopticchurch.com/

Coptic Reading Room for the church in Japan

FileViewGoFavoritesProgramsHelpKizugawa church Ome church

Address

copt-assist.vercel.app

Project

Developed by
So Miyagawa

For the church
in Japan

In development

Coptic Reading Room, public interface observed 7 November 2026

Coptic in Japan

33 / 40

⏱ 32:33–33:34 · Clicks: 2 ｜日本コプト正教会のために宮川が開発中と紹介。既存の公式Coptic Readerとの同一性や、教会の公式承認・導入済みを主張しない。

I am developing Coptic Reading Room for the Coptic Orthodox community in Japan. Its public address is copt-assist.vercel.app. The aim is to help people follow and understand texts used in worship and study, including prayers, biblical passages and saints’ lives. Coptic, Arabic, English and Japanese can be read together. The interface also offers Japanese, English and Arabic navigation. This is an independent project in development. Its Japanese and English translations are labelled as new and unreviewed, so the next stage includes checking the text and terminology with knowledgeable readers. The purpose is practical participation: being able to locate a passage, follow it, and understand the words being read.

Sources:
Project purpose: So Miyagawa, supplied in this presentation request. Public interface observed 7 November 2026.
https://copt-assist.vercel.app/
https://copt-assist.vercel.app/#coverage

Read a passage, then examine its words

FileViewGoFavoritesProgramsHelp

Follow	Understand	Return
Four languages	Word analysis	Search / bookmarks

↓ live demo

Coptic Reading Room, public interface observed 7 November 2026

Coptic in Japan

34 / 40

READING.AVI - Media Player

0:00 / 0:00

Screen recording

⏱ 33:34–34:38 · Clicks: 4 ｜四言語の位置を指す。本文の語を選択→語彙・文法→出典という順で実演。音声・楽譜の内蔵機能は未実装なので紹介しない。

Last step opens a recorded demo (READROOM.AVI: Basil, four columns → word analysis, ~21 s, muted, loops). Talk over it, then press → for the next slide.

This is the Prayer of Preparation in the Liturgy of St Basil. The four versions are aligned by passage. Readers can choose four columns, a two-by-two arrangement or stacked text, and adjust the size. Word panels provide Coptic segmentation, lemmas and parts of speech; Arabic analysis and pronunciation aids are also available. The app offers full-text search, bookmarks, notes and role filters. Source links and explicit gaps let readers return to the underlying texts. Coverage varies by work, and proposed analyses still need contextual checking. I would demonstrate one passage and one word, then return to the source. That is a small, repeatable reading activity that a teacher or learner can use.

Sources:
Coptic Reading Room, public interface and Features and Content Coverage, observed 7 November 2026.
https://copt-assist.vercel.app/#read/basil-0200
https://copt-assist.vercel.app/#coverage

Different tools, different evidence

FileViewGoFavoritesProgramsHelp

Function	Evidence to inspect
Dictionary	Entry and grammatical labels
Next-word suggestion	Corpus examples and counts
LLM-RAG	Retrieved text and generated answer

Decision

The user chooses what fits the sentence.

Observed SINUHE and THOTH AI interfaces

Evaluation

35 / 40

⏱ 34:38–35:35 · Clicks: 4 ｜三つの機能の根拠の違いを示す。

The table distinguishes the evidence behind the different functions. A dictionary lookup gives an entry and grammatical information. A next-word suggestion provides corpus examples and counts. An LLM-RAG system generates a longer answer using retrieved material. The amount and kind of checking differ in each case. For a dictionary entry, we choose the sense and check the variety. For a continuation, we check syntax and intended meaning. For a generated translation, we must also inspect whether the answer adds unsupported content. These tools can work together, but the user still makes the decision about what belongs in the text.

Sources:
https://sinuhe.vercel.app/
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Besa and the Psalms

FileViewGoFavoritesProgramsHelpWhite Monastery

Besa 3:1:2

ϯⲛⲁϩⲱⲙ ⲙⲡⲁⲧⲟⲟⲩⲉ ⲉϫⲛϯⲇⲟⲩⲙⲁⲓⲁ

“I will trample with my shoe upon Edom”

Psalm 59:10 LXX

ϯⲛⲁⲥⲟⲟⲩⲧⲛ ⲙⲡⲁⲧⲟⲟⲩⲉ ⲉϫⲛϯⲇⲟⲩⲙⲁⲓⲁ

“I will extend my shoe over Edom”

Beshara & Miyagawa (2026), slide 40. Spacing simplified

Evaluation

36 / 40

⏱ 35:35–36:36 · Clicks: 4 ｜共通部分と二つの動詞を指す。

Here is another example from the Göttingen presentation, this time involving text reuse. In Besa, the speaker says that he will trample with his shoe upon Edom. The Psalm passage contains the closely related expression about extending the shoe over Edom. The shared wording makes this a passage we can compare directly, while the changed verb deserves attention. A retrieval tool can help us find such candidates across a large collection. Establishing the relation still requires a reading of both contexts. This example shows the value of an inspectable candidate. It does not by itself provide a precision or recall estimate for a detection system.

Map (optional): White Monastery — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Meroitic: text restoration

FileViewGoFavoritesProgramsHelpMeroë

216

masked
restorations

Candidate rank · Agreement

Top 1

51%

Top 5

68%

Status

Preliminary agreement with editorial answers.

Meroitic restoration project, preliminary results

Evaluation

37 / 40

⏱ 36:36–37:32 · Clicks: 3 ｜ 51%と68%は編集者の復元案との一致率。

Meroitic gives us a different prediction task. The preliminary project masks 216 editorial restorations and ranks possible completions. The reported agreement with the editorial answer is fifty-one percent for the first candidate and sixty-eight percent within the first five. The point of comparison is the editorial restoration. That answer may itself be uncertain, and agreement does not establish a correct interpretation of the language. In a follow-up evaluation, I would examine the data split and similarities among formulaic inscriptions. Predicting a sequence can help prioritize possibilities for discussion, while decipherment requires a much broader linguistic and historical argument.

Map (optional): Meroë — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Translation results at a glance

FileViewGoFavoritesProgramsHelpYoron Island Xiulin (Truku)Wadi el-Natrun

Translation	Without	With RAG
Middle Egyptian	32.46	35.43
Bohairic	31.40	40.63
Yoron	35.7	45.6
Truku	4.35	7.47

Reading the table

Different tests. Compare within each row.

Miyagawa (2025–2026), BLEU / SacreBLEU as reported

Evaluation

38 / 40

⏱ 37:32–38:29 · Clicks: 5 ｜同じ行の中で比較するよう促す。

This table brings together the translation results. Each row shows a higher score for the retrieval-assisted system in that particular study. Please read across each row. The test material, translation direction, model configuration and metric implementation differ between rows. We therefore cannot use these numbers to conclude, for example, that one language is easier or that a particular corpus property explains the differences. The Bohairic comparison also involves different deployed interfaces. What the results support is continued investigation within each task. Controlled comparisons could then test retrieval units, dictionary coverage and grammatical annotation while keeping the other conditions fixed.

Map (optional): Yoron Island · Xiulin Township · Wadi el-Natrun — click the blue place name at the top right of the window for a map window; Ctrl+click opens Google Maps in a new tab.

Evaluating use

FileViewGoFavoritesProgramsHelp

Address

Learning library — prototype

Activity	Evaluation
Translate	Error analysis
Learn	Recall
Write	Revision

What can people do with the tools?

Beshara & Miyagawa (2026), learning library prototype

Evaluation

39 / 40

⏱ 38:29–39:26 · Clicks: 4 ｜学習・作文の実践へ話を戻す。

I would like to end with use. The image shows a learning-library prototype from our joint presentation. For translation, evaluation can focus on the errors a specialist or speaker finds. For learning, we can examine whether the user understands a form and retains a pattern. For writing, we can observe revision and readers’ comprehension. Over a longer period, revival research can ask whether these activities become regular practices. The historical discussion helps us keep those aims distinct. My next question is therefore what people can do with the tools, and which forms of evidence make their decisions more reliable.

Sources:
https://docs.google.com/presentation/d/1xzMMo48E1pU7pWmHdvhDr36xtRFC0dasRH-Furx8sGY/edit

Thank you

FileViewGoFavoritesProgramsHelpUniv. of Tsukuba The Met

ϯϣⲉⲡϩⲙⲟⲧ

So Miyagawa

miyagawa.so.kb@u.tsukuba.ac.jp

Appendix
(Q&A)

Coptic textile

The Met, 89.18.95 · public domain

Thank you — questions welcome

Thank you

40 / 40

Appendix: Yoron results

FileViewGoFavoritesProgramsHelpYoron Island Amami Islands

System	BLEU	chrF	METEOR	ROUGE-L	TER ↓
Yunnufutuba + RAG	45.6	49.8	67.3	68.4	38.8
Claude 4.5 Opus	37.6	40.5	58.4	59.8	45.6
Gemini 3.0 Pro Preview	35.7	38.8	55.2	57.4	48.4
GPT-5.2 Thinking	22.6	25.1	35.2	40.2	62.6
DeepSeek-V3.2	17.7	20.5	28.6	33.1	71.5
Kimi 2.5 Instant	15.5	16.5	23.6	29.8	75.4
Grok 4.1	7.3	12.3	16.0	20.4	90.4

14 sentences. Historical model labels and reported scores are retained.

Miyagawa (2026a), Table 3

Appendix

41 · appx

Appendix: Truku pilot results

FileViewGoFavoritesProgramsHelpXiulin (Truku)Wanrong Zhuoxi

System	BLEU	chrF	METEOR	ROUGE-L	TER ↓
Claude Fable 5 Max	11.20	41.52	40.57	44.37	80.75
DeepSeek-V4-Pro-0813	7.62	35.07	33.55	41.25	75.94
Kimi-K3	7.49	40.83	38.92	40.42	77.01
Truku Translator AI + RAG	7.47	37.92	36.72	40.99	76.47
GPT 5.6 Sol Ultra	7.03	40.28	35.15	41.84	82.89
GLM-5.3	5.58	38.39	35.15	37.31	82.89
Gemini 3.7 Flash	4.35	36.92	30.04	35.73	80.75
DeepSeek	1.73	31.87	22.64	24.52	103.21
Qwen 3.7 Plus	1.64	21.32	16.69	28.29	86.10
Gemma-4-31B-it	0.30	15.52	4.09	7.04	111.23

Gemini 3.1 Pro and MiniMax-M3 declined. Six paragraphs, one reference.

Miyagawa, Tang, Sato & Koizumi (2026), Truku pilot

Appendix

42 · appx

Appendix: Bohairic translation results

FileViewGoFavoritesProgramsHelpWhite Monastery Wadi el-Natrun

System	BLEU	chrF	R-1	R-2	R-L	TER ↓	METEOR
THOTH AI (Gemini)	40.63	63.84	76.85	48.76	69.95	44.12	70.35
Gemini 3 Pro Preview	31.40	60.20	70.87	42.16	66.02	50.00	66.78
Claude Opus 4.5	29.97	56.91	70.14	40.19	63.51	50.98	64.70
GPT 5.2 Pro	16.74	50.80	57.27	31.19	55.45	67.65	47.75
Grok 4.1	16.32	49.36	59.33	29.95	53.59	64.71	48.33
Kimi K2	6.64	40.88	48.87	15.53	32.58	90.20	36.66
DeepSeek V3	4.76	33.69	39.79	10.58	27.23	83.33	24.08
Qwen3 Max	3.38	30.55	37.44	3.69	19.18	102.94	22.20

One passage-level segment; one retained output for each deployed system.

Miyagawa (2026c), Table 2

Appendix

43 · appx

Appendix: Middle Egyptian results

FileViewGoFavoritesProgramsHelpElephantine

System	SacreBLEU	chrF	METEOR	ROUGE-1	ROUGE-L
THOTH AI + RAG	35.43	61.05	.650	.730	.680
Claude 3.5 Sonnet	32.46	58.06	.640	.717	.652
Gemini 2.0 Pro	28.77	59.94	.581	.714	.681
Gemini 2.0 Flash	25.59	59.30	.571	.697	.664
DeepSeek R1	22.16	53.74	.510	.675	.587
GPT o1 Pro	21.60	53.17	.509	.625	.534
GPT 4o	19.62	51.11	.417	.581	.484
GPT o3-mini-high	3.41	35.15	.249	.375	.233

METEOR and ROUGE retain the original 0–1 scale in this table.

Miyagawa (2025), NLP4DH, Table 2

Appendix

44 · appx

Appendix: resources and retrieval settings

FileViewGoFavoritesProgramsHelpYoron Island Xiulin (Truku)

System	Retrieval collection	Configuration reported
THOTH AI (2025)	Comprehensive Coptic Lexicon with Egyptian etymologies	Dify; Claude 3.5 Sonnet
THOTH AI 2.0	CCL, Coptic SCRIPTORIUM and Faulkner	Vector search; top 3; Gemini 3 Pro Preview
Yunnufutuba	9,007 Yoron–Japanese bunsetsu pairs	text-embedding-3-large; top 6
Truku translator	Aligned Bible verses and dictionary entries	text-embedding-3-large; top 6

Model versions describe the reported experiments; current applications may differ.

Miyagawa (2025–2026), system configurations

Appendix

45 · appx

Appendix: the scale of available resources

FileViewGoFavoritesProgramsHelpYoron Island Xiulin (Truku)Hokkaido (Ainu)

Resource	Coptic	Ainu	Truku	Yoron
Corpus tokens	2.38M	45K	116K	—
Dictionary headwords	11,263	3,467	9,057	about 1K
Parallel material	multiple resources	about 220K pairs	Bible translation	9K bunsetsu pairs
Speech hours	limited	about 39	30–50	—

Counts describe different resources. Their size alone cannot explain differences in translation scores.

Miyagawa (2026b), JSAI

Appendix

46 · appx

Appendix: related work on learning from supplied resources

FileViewGoFavoritesProgramsHelpKaras Is. (Kalamang)Guangxi (Zhuang)Taiwan (Formosan)Kagoshima

Study	Resource or question
Tanzer et al. (2024), MTOB	Translation with a grammar book for Kalamang
Aycock et al. (2025)	Contributions of parallel examples and grammatical descriptions
Zhang et al. (2024), DiPMT++	Dictionaries and parallel examples for an unseen language
Zhang et al. (2024), LingoLLM	Linguistic descriptions and tools in the model context
Zheng et al. (2024)	Bilingual lexical resources for Formosan languages
Sakai (2024)	Kagoshima dialect generation with supplied language resources

See bibliographic references

Appendix

47 · appx

Appendix: what the last-speaker question can tell us

FileViewGoFavoritesProgramsHelpEgypt

Evidence	What it can show	What it cannot settle alone
A dated manuscript	Copying, reading or learned composition	Native acquisition in the writer’s household
Arabic grammar of Coptic	Deliberate teaching to Arabic readers	Identical competence in every region
Report of a late speaker	A claim requiring historical scrutiny	Continuous community transmission
Present-day new writing	An active learned or revival practice	The scale of everyday use

The chronology changes with the domain and the evidence we count.

Methodological distinction developed from Richter (2009; 2023) and Teepe (2025)

Appendix

48 · appx

References: Coptic history and revival

FileViewGoFavoritesProgramsHelp

Richter, T. S. 2009. Greek, Coptic and the “language of the Hijra”. In From Hellenism to Islam, 401–446. doi:10.1017/CBO9780511641992.019

Richter, T. S. 2023. Coptic. UCLA Encyclopedia of Egyptology. escholarship.org/uc/item/22r6s881

Teepe, R. 2025. The Death and Afterlife of the Coptic Language in Mamluk Egypt. PhD dissertation, Yale University. elischolar.library.yale.edu/gsas_dissertations/1663/

Miyokawa, H. 2016. Pharaonism and the Revival of the Coptic Language among Early Twentieth-Century Coptic Christians. Bulletin of the Society for Near Eastern Studies in Japan 58(2): 184–195. doi:10.5356/jorient.58.2_184

AUC Libraries. Coptic Studies: Iryan Moftah Collection and Ragheb Moftah Archives. library.aucegypt.edu/rare-books-special-collections/collections/coptic-studies

References.txt

Appendix

49 · appx

References: project studies and language resources

FileViewGoFavoritesProgramsHelp

Beshara, M. & S. Miyagawa. 2026. Assistance of Analysis, Translation, and Composition in Coptic Language through Artificial Intelligence: THOTH AI 2.0 and Coptic Composer AI. 13th ICCS, Göttingen, 31 July.

Miyagawa, S. 2025. RAG-Enhanced Neural Machine Translation of Ancient Egyptian Text: A Case Study of THOTH AI. NLP4DH 2025, 33–40.

Miyagawa, S. 2026a. Application of LLM-RAG to Machine Translation of Yoron Dialect in Amami Islands. IEICE Technical Report, NLC. 2026b. Neural Machine Translation for Endangered Languages Using LLM-RAG. JSAI, 4I4-OS-17a-01.

Miyagawa, S. 2026c. Retrieval-Augmented Translation for Bohairic Coptic: A Pilot Case Study. LaTeLL 2026. Miyagawa, Tang, Sato & Koizumi. 2026. Improving Machine Translation for Truku Seediq with RAG. NINJAL symposium.

Burns et al. 2020. Comprehensive Coptic Lexicon v1.2. doi:10.17169/refubium-27566. Coptic SCRIPTORIUM: copticscriptorium.org. SINUHE: sinuhe.vercel.app.

References.txt

Appendix

50 · appx

References: texts and comparative NLP work

FileViewGoFavoritesProgramsHelp

Kiku, C. 1985. 与論のしまがたり. Haru Shobō. Kiku, H. 2017. 与論の言葉で話そう. Nanpō Shinsha. Kibe, N. 2019. 奄美方言のエビデンシャリティ. 国際日本学研究報告 5: 33–46.

Leipoldt, J. & W. E. Crum. 1906. Sinuthii archimandritae vita et opera omnia I. CSCO 41. Bell, D. N. 1983. The Life of Shenoute by Besa. Cistercian Publications. Nederhof, M.-J. 2023. Digital editions of Middle Egyptian texts.

Lewis et al. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS. Tanzer et al. 2024. A Benchmark for Learning to Translate a New Language from One Grammar Book. ICLR.

Aycock et al. 2025. Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? ICLR. Zhang, C. et al. 2024. Teaching Large Language Models an Unseen Language on the Fly. Findings of ACL.

Zhang, K. et al. 2024. Hire a Linguist! Findings of ACL. Zheng et al. 2024. Improving Low-Resource MT for Formosan Languages Using Bilingual Lexical Resources. Findings of ACL. Sakai, M. 2024. 生成AIを用いた鹿児島方言生成. NLP2024.

References.txt

Appendix

51 · appx

Appendix: places on Google Maps

FileViewGoFavoritesProgramsHelp

Click a place for a map window; Ctrl+click opens Google Maps in a new tab. Nazlet al-Fallahin: village position from GeoNames.

Google Maps; village position for Nazlet al-Fallahin: GeoNames

Appendix

52 · appx

Appendix: links — tools, data and languages

FileViewGoFavoritesProgramsHelp

Language	Glottolog	ISO 639-3	Ethnologue	Wikipedia
Yoron [yox]	Glottolog	ISO 639-3	Ethnologue	Wikipedia
Seediq (Truku) [trv]	Glottolog	ISO 639-3	Ethnologue	Wikipedia
Egyptian [egy]	Glottolog	ISO 639-3	—	Wikipedia
Coptic [cop]	Glottolog	ISO 639-3	Ethnologue	Wikipedia
Meroitic [xmr]	Glottolog	ISO 639-3	—	Wikipedia
Ainu [ain]	Glottolog	ISO 639-3	Ethnologue	Wikipedia
Kalamang [kgv]	Glottolog	ISO 639-3	Ethnologue	Wikipedia

Links checked 8 November 2026

Appendix

53 · appx

Appendix: links — publications and talks

FileViewGoFavoritesProgramsHelp

ACL Anthology, arXiv, DOI, J-STAGE, IEICE, NDL Search, CiNii, WorldCat

Appendix

54 · appx

Appendix: links — background reading and contact

FileViewGoFavoritesProgramsHelp

Wikipedia and author pages

Appendix

55 · appx

Appendix: links — church, news and project pages

FileViewGoFavoritesProgramsHelp

Official church site, Christian Today, Kirishin, AMOR, Al Manassa, project pages

Appendix

56 · appx

It’s now safe to close
this presentation.

ϯϣⲉⲡϩⲙⲟⲧ