1
Enhancing Digital Cultural Heritage
with Computational Methods
and Human Participation
AI-assisted metadata enrichment • Human-in-the-loop crowdsourcing • Reproducible workflows
PILOT
EUROPEANA
CROWDHERITAGE
Dr Katerina Zourou, Mariana Ziku
Web2Learn
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
2
1. Problem framing
From visual availability to semantic discoverability
“
Digital collections may be visible as images, but remain hard to find when the right descriptive words are missing.
Visual access
Users can see images when they reach the record, but search depends on existing metadata.
Semantic access
Searchable tags describe figures, objects, scenes, places, and motifs.
Research question
How can computational methods and human participation jointly enrich cultural heritage metadata at scale?
Take-away
AI adds scale.
Humans add context, trust, and cultural interpretation.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
3
2. The pilot at a glance
A case study in AI-assisted and participatory metadata enrichment
Dimension
Operational choice
Project
AISTER Erasmus+ pilot
Collection
Krovets Ukrainian folk paintings and icons
Platforms
Europeana + CrowdHeritage
Methods
NLP, computer vision, crowdsourcing, ethical screening
Outputs
Open notebooks, CSV/JSON-LD datasets, reusable workflow
312
gallery artefacts
70
contributors
5
workshops
8
notebooks
5,946
annotations
51,952
Activity points
“
The pilot treats metadata enrichment as a socio-technical process: computationally assisted, publicly validated, and openly documented.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
4
3. Corpus and data provenance
Following the collection from source institution to participatory platform
Krovets Online Museum
source
institution
MUSEU
aggregator
Europeana
access platform �& APIs
CrowdHeritage
validation platform
Open repositories
GitHub + Zenodo
Corpus
312 folk paintings and icons, drawn from a wider ethnographic collection of 3,840 artefacts.
Content
Rural life, religious imagery, portraits, landscapes, motifs, and material culture.
Data access
A Europeana gallery and API endpoint streamlined the retrieval process.
Focus
Discoverability through new description tags.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
5
4. Methodological architecture
A four-step human-in-the-loop workflow
1
Metadata → NLP
Generate candidate tags from textual metadata
2
Images →
Computer Vision
Caption images and extract visual tags
3
Crowd validation
Validate, reject, and add annotations
4
Ethical screening
Check terms against DE-BIAS vocabulary
Design principle
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
6
5. Step 1 – Natural Language Processing from textual metadata
Turning catalogue descriptions into candidate tags
Input
Title, description, subject, creator, item type, place, and date metadata.
Processing
Rule-based heuristics and spaCy Named Entity Recognition.
Output
Candidate tags for figures, objects, and scenes.
893
text-based annotations
Tag family
Examples of descriptive value
Figures
woman, farmer, child, historical figure
Objects
horse, candle, chair, icon
Scenes
village, celebration, church, rural landscape
“
NLP provides a structured first pass over existing metadata, revealing descriptive signals already present in catalogue records.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
7
6. Step 2 - Computer vision and image captioning
Generating tags from visual content, not only metadata
Image retrieval
Europeana gallery artefacts downloaded for processing
Captioning
Qwen visual-language models generate image descriptions
Tag extraction
Captions are transformed into structured candidate tags
4,581
raw image-based annotations
3,927
combined after deduplication
What computer vision adds
Visual tags can describe attire, backgrounds, objects, damage, scenery, text, and figures that may not appear in the original metadata.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
8
7. Step 3 - Human-in-the-loop crowdsourcing
Validation as a participatory quality-assurance layer
CrowdHeritage task model
Upvote
confirm accurate tag
Downvote
reject misleading tag
Add tag
contribute human annotation
Votes by workshop/event
W1
6,416
W2
12
W3
11,084
W4
33,361
W5
1,079
green = upvotes�red = downvotes
value = total votes
70
participants
5
workshops
51,952
Activity points
“
Human validation converts AI outputs into accountable, more trustworthy cultural data.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
9
8. Semantic modelling and reproducibility
Making the workflow reproducible, reusable, and citable
Europeana APIs
retrieve records and metadata
Jupyter / Colab
run executable notebooks
JSON-LD / W3C Web Annotation
model annotations for platform ingestion
GitHub
version control and documentation
Zenodo
long-term preservation and DOI of open dataset
Web Annotation data model example for ingestion in the CrowdHeritage platform
{
"type": "Annotation",
"creator": "AI model or human",
"body": { "value": "tag" },
"target": "Europeana item",
"review": { "upvotes": n, "downvotes": n }
}
“
Reproducibility becomes: methods, data, code, and outputs remain traceable, reproducible.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
10
9. Results and data insights
From AI outputs to human-validated descriptive metadata
311
artefacts in final snapshot
5,946
total annotations
3,917
software annotations
2,029
human annotations
48,599
upvotes
3,353
downvotes
Annotation composition
Software 66%
Human 34%
Accepted more often
Rejected more often
icon
cracks
painting
wear
man / woman
damage
trees
small object / staff
Source: AISTER pilot documentation and supplied Web2Learn/AISTER materials
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
11
10. Ethical assessment and data quality
Why validation is not the end of the workflow
Quality-control sequence
AI-generated tag
Human-approved tag
DE-BIAS screening
Manual revision
Observed intervention
Original term
Ethically revised term
Status
slave
enslaved person
accepted during validation; revised after screening
Take-away
Human approval can still reproduce problematic terminology. Ethical screening adds a further layer of accountability.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
12
11. Contributions and remarks
What this pilot demonstrates for digital cultural heritage research
Methodological
A concrete workflow combining AI (NLP, computer vision), crowdsourcing validation, and ethical assessment.
Participatory science
A crowdsourcing campaign for annotation validation of Ukrainian folk art, with metrics-based insights.
Infrastructural
An open technical stack: APIs, notebooks, JSON-LD, GitHub, Zenodo.
Output
Open datasets with enriched metadata and user metrics.
Cultural heritage discoverability is not only a technical problem. �Ιt is a participatory, interpretive, and ethical process.
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage
13
Thank you
Access the AISTER pilot
AISTER HITL Crowdsourcing Pilot | Digital Cultural Heritage