Building computational workflows for human rights archives

Authors

In brief

Software built for human rights archives faces a design paradox that is rarely discussed in the literature: the data that drives its development cannot be made public. Drawing on two ongoing academic projects, Memorias in Uruguay and NuestraMemorIA in Chile, the authors describe a two-layer architecture for applying natural language processing to declassified military and government archives from the Southern Cone dictatorships of the 1970s, and set out the methodological decisions that data sensitivity imposes at every step.

“At every step, decisions on access, storage, and inference become ethical choices.”

What the article covers

• Layer one, document recovery: optical character recognition on mimeographed, hand-annotated and badly degraded pages, followed by named entity recognition and information extraction.
• Local inference as a legal and ethical requirement rather than a cost decision, and sovereign retrieval as the principle that governs every infrastructure choice that follows from it.
• Layer two, search and aggregate analysis: LUZ and Amalia in the Uruguayan project, a knowledge-graph chatbot grounded in explicit graph evidence in the Chilean one.
• The evaluation problem when neither documents nor annotations can be released: synthetic corpora, public subsets read as upper bounds, and validation by historians and archivists in the loop.
• A worked example, from a keyword search through table extraction on a detention list to a partial detention trajectory recovered by record linkage.
• Sensitivity treated as a spectrum rather than a binary, with access control at document level and not at collection level.

How to cite

Elina Gómez, Lorena Etcheverry and Jocelyn Dunstan. “Building computational workflows for human rights archives.” vISIon: The ISI Magazine Vol. 1 No. 2, June 2026, pp. 30–34.