Metadata doesn’t just describe information — it decides who finds it, how they find it, and whether it makes sense when they do.

I had the amazing opportunity to work with the Digital Theological Library (DTL) through OCLC, and oh, what a delight it was. Initially I was creating multilingual collections, but in looking at the data available, I kindly asked if I could design a different project. Given my interest in metadata — and multilingual metadata at that — I created a project I called Semantic Enrichment of Cross-Institutional, Multilingual Metadata. It started as a technical challenge: scraping and analyzing dissertation metadata across three different institutions in three different languages. But it quickly turned into something bigger: a study of how knowledge travels — or doesn’t — across borders, vocabularies, and cultures.

For example, Santería in Cuba is not the same as Candomblé in Brazil, even if both are Afro-diasporic religions. And yet, a searcher looking in only one language might never know the other exists. My goal was to make those relationships visible without erasing the differences that matter.

Each institution in this study — University of Maryland (U.S.), Universidad de los Andes (Colombia), and Universidade Presbiteriana Mackenzie (Brazil) — catalogs its dissertations in its own language and according to its own internal standards. There was no shared vocabulary, no cohesive structure, and almost no cross-lingual subject mapping. Even though all three institutions published work on overlapping topics, those connections were practically invisible in the metadata.

Building the Controlled Vocabulary

To address the gaps I was seeing, I took a hands-on approach — treating this not just as a research question, but as a full-stack metadata project, top to bottom. I first built a trilingual controlled vocabulary by hand that would be used to scrape the data and generate the custom dataset. I focused on terms related to Afro-diasporic religions, Indigenous communities, colonial legacies, and spiritual practices; pulling from LCSH where it helped, but also incorporating regionally specific terms I encountered during scraping and abstract analysis. This vocabulary became the backbone for everything that followed.

Scraping and Normalization

Each institutional repository had its own quirks: different site structures, different field names, different levels of accessibility — mostly DublinCore. So I wrote a custom web scraper in Python for each one, targeting the dissertation metadata. Once I had it all, I built a normalization pipeline to standardize the fields across Spanish-, Portuguese-, and English-language sources.

The Language Work: NER and Semantic Embeddings

I focused primarily on the subject and abstract elements of the records, as these carry the richest narrative content. Using spaCy’s NLP models for Spanish, Portuguese, and English, I applied Named Entity Recognition (NER) to pull out key noun phrases and proper names from each abstract. This surfaced important concepts missing from the official subject fields — terms like bata drums, regalías, or quilombolas that carry deep cultural and political significance.

I then used a multilingual transformer model to embed all subject terms into the same vector space, calculated similarity scores, and built clusters — essentially creating a semantic map that aligned, say, “derechos humanos,” “direitos humanos,” and “human rights” not just as translations, but as conceptually equivalent terms that might now surface in the same search.

Finally, I ran TF-IDF vectorization and KMeans clustering on the enriched dataset. Patterns around ritual, identity, and postcolonial resistance emerged that were otherwise buried by inconsistent or sparse metadata. Together, these steps helped transform isolated metadata fields into a network of multilingual, semantically linked concepts.

What Emerged

This project turned up more than just better metadata. It revealed how institutional and linguistic contexts shape research focus. While there is no evidence of subject term translation bias, there is a coverage gap.

The Enriched Records

At the end of the project, I produced enriched records that included both original and semantically expanded subject terms. Here is a sample record and the schema it follows.

{
  "dc.title": "Afro-Cuban Bata Drum Aesthetics",
  "dc.title.alternative": "Lucumí Music and Trance Rituals in Cuba",
  "dc.contributor.author": ["Martinez, Isabella", "Dr. Ana Rodríguez"],
  "dc.date.issued": "2022-08-10",
  "dc.subject": [
    "Music",
    "Anthropology, Cultural",
    "Education, Music"
  ],
  "dc.subject.enriched": [
    "Music",
    "Anthropology, Cultural",
    "Education, Music",
    "Lucumí religion",
    "Santería",
    "Yoruba religion",
    "bata drums",
    "possession trance"
  ],
  "dc.description.abstract": "This study explores the Lucumí religion in Cuba...",
  "dc.publisher": "University of Maryland",
  "dc.rights": "Creative Commons Attribution 4.0 International",
  "dc.language": "en"
}

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "title": "Multilingual Dissertation Metadata",
  "type": "object",
  "properties": {
    "dc.title":               { "type": "string" },
    "dc.title.alternative":   { "type": "string" },
    "dc.contributor.author": { "type": "string" },
    "dc.date.issued":         { "type": "string", "format": "date" },
    "dc.subject":            { "type": "array", "items": { "type": "string" } },
    "dc.subject.enriched":   { "type": "array", "items": { "type": "string" } },
    "dc.description.abstract": { "type": "string" },
    "dc.publisher":          { "type": "string" },
    "dc.identifier.uri":     { "type": "string", "format": "uri" },
    "dc.language":           { "type": "string" },
    "dc.source":             { "type": "string" }
  },
  "required": [
    "dc.title", "dc.subject", "dc.description.abstract",
    "dc.identifier.uri", "dc.language", "dc.source"
  ]
}