                 Entities and experiments. Corpus exploration, proto indexing and enrichment workflows

                                     Digital Scholarly Editions "Arcipelago Ceresa"

            Named Entities in Digital Editions. Between Structured Databases and Context-Specific Annotation

                                      Open Editions Workshop, University of Zurich

                                                       07.03.2026

                                   Levyn Bürki, Data Science Lab, University of Bern

                                  Peter Dängeli, Data Science Lab, University of Bern

                                 Edition workflow: General approach

                 ██ Corpus definition

                 ██ Image digitisation (Swiss National Library)

                 ██ IIIF upload

                 ██ Transkribus (raw transcriptions)

                 ▓▓▓ Generation of IIIF manifests (as sequences of canvasses/images

                 per document)

                 ▓▓▓ IIIF manifest-based upload to Transkribus

                 ▓▓▓ Transcription (primarily automated for print/typescripts,

                 manual for manuscripts)

                 ▓▓▓ Document export (incl. transformation to project data

                 structure)

                 ██ Transcription and annotation in oXygen XML editor (with project

                 framework)

                 ██ Web app development (SvelteKit, CETEIcean)

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 2 / 12

                                 Edition workflow: General approach

                 ██ Corpus definition

                 ██ Image digitisation (Swiss National Library)

                 ██ IIIF upload

                 ██ Transkribus (raw transcriptions)

                 ▓▓▓ Generation of IIIF manifests (as sequences of canvasses/images

                 per document)

                 ▓▓▓ IIIF manifest-based upload to Transkribus

                 ▓▓▓ Transcription (primarily automated for print/typescripts,

                 manual for manuscripts)

                                                                                         Named entity

                 ▓▓▓ Document export (incl. transformation to project data               detection

                 structure)                                                              and

                                                                                         identificati

on

                 ██ Transcription and annotation in oXygen XML editor (with project

                 framework)

                 ██ Web app development (SvelteKit, CETEIcean)

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 3 / 12

                                  Edition workflow: Named entities

                 Goal: facilitate manual tagging by creating an automatically

                 compiled list with possible entries (including authority

                 references)

                 Approach:

                    •  use an LLM to evaluate the raw Transkribus output of each

                       transcribed document and

                    •  detect entities,

                    •  try to link them to one or more authority records, and

                    •  deduplicate the resulting list

                    •  in order to offer the entries in the linking utility in

                       oXygen

                 We call this a "proto index", an imperfect index of entities that

                 likely occur in the corpus.

                 It is meant to be a shortcut for the editors that saves them

                 manual querying of authority databases.

                 The actual entity tagging is done by the editors.

                 State of work: experimental, explorative (but not too far from

                 production-ready)

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 4 / 12

                                 Demo

                 ██ Step 1                                         ██ Step 2

                 Feed documents to LLM and ask to                  Take output of step 1 and enrich it

                 detect entities, then identify them               with Wikidata information, then

                 using Wikidata knowledge.                         generate a spreadsheet (csv).

                 Using a slim local helper service                 Simple Python pipeline executing

                 that acts like a plug‑in the model                SPARQL queries (no LLM used).

                 can call (via the Model Context

                 Protocol, a simple way for tools to

                 talk to each other).

                 This allows to search (e.g.)

                 Wikidata directly (quickly, in

                 memory) without extra servers or

                 manual wiring.

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 5 / 12

                                                          Demo

     ██ Step 1

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 6 / 12

                                                          Demo

     ██ Step 1

     Feed documents to LLM and ask to detect entities, then identify them using Wikidata knowledge.

                ENTITY_CLASSES = [

                    "persons", "places", "institutions", "publishers", "works", "events", "citations"

                ENTITY_SEARCH_SYSTEM_PROMPT = f"""

                You are a semantic annotation assistant for a Digital Humanities project working with

                historical writings (Italian).

                Your task: read the text below and extract ALL named entities you can find, grouped by

                category.

                Return a JSON object with the following top-level keys, each mapping to an array of

                exact surface forms as they appear in the text:

                {json.dumps({category: [] for category in ENTITY_CLASSES}, indent=2)}

                Rules:

                - Use the exact string as it appears in the source text (do not normalize or modernize)

                - If a category has no entries, return an empty array

"""

                ENTITY_LINKING_SYSTEM_PROMPT = """

                You are an entity linking assistant for a Digital Humanities project working with

                historical writings (Italian).

                Your task: link the given entity to a knowledge base (e.g. Wikidata) using the provided

                context for disambiguation.

                Return a JSON object with a 'candidates' key containing a list of matches, each with a

                'qid' and a 'confidence' of high, medium, or low. If no candidates are found, return an

                empty list.

"""

                TEXT_SUMMARIZATION_SYSTEM_PROMPT = """

  Arcipelago Ceresa are  Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 7 / 12

                historical writings (Italian).

                                     pip install -r requirements.txt

                                     python main.py > output.txt

                          For executed code see: https://asciinema.org/a/CvUyVktwWlx02NTE?t=43

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 8 / 12

                pip install -r requirements.txt

                python -m main --input ../../demo/step1/output.txt --output entities_enriched.csv --log

                pipeline.log

                         For executed code see: https://asciinema.org/a/CvUyVktwWlx02NTE?t=165

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                 9 / 12

                                     csvlens demo/step2/entities_enriched.csv

                         For executed code see: https://asciinema.org/a/CvUyVktwWlx02NTE?t=194

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                10 / 12

                 ██ Result: facilitated manual

                 tagging based the automatically

                 compiled list

                 The generated entries are used to

                 populate the entity spreadsheet(s)

                 of the project (with manual checks).

                 The oXygen framework queries the

                 spreadsheet and offers the entities     ——————————————————————————————————————————————

                 for comfortable linking.

                 ————————————————————————————————————

                 ██ Next steps and further

                 considerations

                 ▓▓▓ Workflow decisions

                 We need to define in what frequency

                 and with what degree of automation

                 this recognition/identification task

                 is executed.

                 One idea is to integrate it into the

                 Transkribus document export. In

                 part, this depends on how fixed the

                 decisions around named entities are

                 (and with them the structure of the

                 spreadsheet).

  Arcipelago Ceresa  Top Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                11 / 12

                                   Technical partner of the Arcipelago Ceresa project

                                        Data Science Lab -- https://dsl.unibe.ch

                                              https://youtu.be/afXUHAUZ4dk

  Arcipelago Ceresa      Named Entities in Digital Editions, Roundtable “Workflows”, 07.03.2026                12 / 12