Designing Recommender Systems for Digital Humanities

Summary of Designing Recommender Systems for Digital Humanities

by Kyle Polich

36m•November 23, 2025
Listen to Episode

Overview of Designing Recommender Systems for Digital Humanities

This episode of Data Skeptic (host Kyle Polich) interviews Florian Atzenhofer about designing recommender systems for digital humanities, with a concrete case study: Monasterium.net — the largest online collection of historical charters. The conversation covers why off‑the‑shelf recommenders are often insufficient for cultural heritage archives, the specific data and user challenges for charters, multimodal similarity approaches (text, image, metadata, named entities), UX choices for different user skill levels, evaluation difficulties and proposed directions, and openness/standards in the DH community.

Key takeaways

  • Digital humanities (DH) archives have unique needs: multimodal items (scans + transcriptions + rich domain metadata), sparse user–item interactions, and diverse, highly invested user groups (experts, novices, genealogists, art historians, linguists).
  • Cold start and extreme sparsity make content‑based and knowledge‑driven approaches essential. Leverage text (transcriptions/Regesta), image features, domain metadata, and named entities (diplomatics).
  • Offer multiple recommendation/exploration modes: guided exploration for novices and controlled similarity/re‑ranking for advanced users. Allow users to combine and weight modalities.
  • Evaluation must go beyond classical accuracy metrics; consider multi‑stakeholder objectives (fairness across collections, research‑path quality, explainability, provenance).
  • Practical pipelines: high‑quality scans → HTR (handwritten text recognition) / OCR → structured XML (Charter Encoding Initiative, inspired by TEI) → multimodal embeddings.
  • Open source, open standards, and federation/aggregation (e.g., Europeana) are desirable and feasible within DH but repositories vary.

Topics discussed

Monasterium.net and the problem context

  • Monasterium.net: 20+ year project hosting millions of medieval and modern charters across Europe (images, transcriptions, abstracts called Regesta, domain metadata).
  • Project DDIP is rebuilding Monasterium and plans to add recommender and semantic search features (new version due by end of year).

Who are the users?

  • Primary users: researchers in history and related domains, but also genealogists, art historians, linguists, students, educators, archivists, and the general public.
  • Users have varied expertise and goals: discovery/serendipity vs. targeted research for publication.

Item definition and data modalities

  • Item = single charter (high‑res image + transcription when available + Regesta abstract + diplomatics metadata such as issuer/recipient/date/location).
  • Data often serialized in CEI (Charter Encoding Initiative) XML; metadata availability depends on digitization pipelines.

Recommender types & UX design

  • Two main UX components:
    • Exploration page (guided discovery): preconfigured profiles for novices, emphasizes serendipity and broader exploration.
    • Controlled similarity/search: build “baskets” of items and re‑rank or expand via similarity (user‑driven weights).
  • Allow weighting of modalities (e.g., prioritize iconography in images, or named entities in metadata).

Technical approaches & multimodality

  • Because of sparse interactions, emphasis on content and knowledge embeddings:
    • Text: TF‑IDF remains useful for historical corpora; skip‑grams, n‑grams, and more modern embedding approaches are also applicable.
    • Images: visual feature embeddings for iconography or illumination patterns.
    • Metadata & named entities: NER adapted to historical language + diplomatics tagging (issuer/receiver roles).
  • Combine modalities into weighted similarity spaces and let users control weighting for personalized exploration.

Evaluation & multi‑stakeholder considerations

  • Classic metrics (RMSE, ranking accuracy) are necessary but insufficient.
  • Proposed “research funnel” stages for evaluation:
    1. Discovery — initial exploration and serendipity.
    2. Interaction — assessing immediate relevance and usefulness.
    3. Integration — deeper comparative use and incorporation into research.
    4. Impact — downstream effects (publications, teaching, editorial use).
  • Metrics suggested: research path quality (subjective, user‑reported), collection representation/fairness (avoid over‑surfacing dominant archives), explainability, and provenance/versioning of recommender outputs.
  • Stakeholders want explainability and control; transparency about software versions/builds (provenance IDs) is desirable for reproducibility and citation.

Openness, standards, and interoperability

  • Many repository software stacks in DH are open source; community favors open standards.
  • Aggregation and federation (national/EU level, e.g., Europeana) present opportunity but require non‑flattened, interoperable data to avoid silos.
  • Recommendation-as-exploration could spread across other cultural heritage repositories.

Notable quotes / insights

  • “We have a very, very cold starting problem…we leverage different modalities of the charter data and embed them in different spaces.”
  • “Charters can be very, very lengthy…Regesta are short summaries that are really important for similarity search and recommendation.”
  • “Stakeholders interpreted serendipity in different ways…some say the best finds are those you wouldn’t expect, found by equal representation.”

Actionable recommendations (for implementers)

  • Start by auditing your item modalities: do you have high‑res images, HTR/OCR transcriptions, Regesta/abstracts, and diplomatics metadata? Prioritize pipelines that add these.
  • Implement multimodal embeddings (text + image + metadata) and allow configurable weighting so users can tailor similarity signals.
  • Provide dual UX modes: a guided exploration interface for novices and focused re‑ranking / basket workflows for advanced researchers.
  • Incorporate explainability and provenance: include human‑readable reasons for recommendations and a version/ID of the recommender build in any exported results.
  • Evaluate across the research funnel: combine conventional ranking metrics with user studies measuring research path quality, serendipity, and collection representation.
  • Engage stakeholders (archivists, domain specialists, educators) early and iteratively — use focus groups/interviews to align priorities and metrics.

Research & next steps (Florian’s PhD)

  • Florian’s PhD focuses on three strands:
    1. Multi‑stakeholder recommender systems for charters (evaluation & fairness).
    2. Information‑seeking behavior of DH user groups.
    3. Multimodal similarity models tailored to cultural heritage data.
  • Upcoming: a paper and PhD symposium participation at the RecSys conference in Prague; continued literature work and experiments in multimodality and evaluation.

Resources and links

  • Monasterium: https://monasterium.net
  • DDIP project: https://ddip.eu
  • Charter Encoding Initiative (CEI) — XML serialization inspired by TEI
  • RecSys conference (where evaluation paper will be presented)

If you want a focused cheat‑sheet for building DH recommenders, follow the Actionable recommendations above: audit modalities, enable multimodality + user weighting, design separate UX flows for novices/experts, track provenance, and evaluate across discovery→impact stages.