New ML pipeline graphs historical actions from Danish archives

In a new working paper on arXiv, researchers Sofus Landor Dam and Johan Heinsen present a pipeline that leverages machine learning to transform unstructured historical sources into structured, action-based graphs. The approach is built on the GRAM framework (Graph of Roles and Actions Model), which treats actions—rather than entities or events—as the fundamental unit of analysis. The pipeline automates the creation of skeletal graphs (auto-GRAMS) from text, enabling researchers to map relationships and patterns of behavior across large datasets without losing the granularity that close reading provides.

The authors test their method on four archival collections from 18th- and 19th-century Denmark, specifically analyzing instances of 'pretending' among runaways and itinerants. By graphing how individuals assumed false identities or roles, the pipeline reveals structural patterns in historical social behavior. The paper emphasizes that auto-GRAMS are not a replacement for manual analysis but a complement—allowing historians to scale their work while preserving the depth needed for social history. This represents a significant step in computational humanities, making large-scale, granular analysis of archival records feasible for the first time.