ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-correction for Historical Documents

The value of digitized historical media archives for computational historical research is now well established, yet an underexplored challenge concerns data management itself: how to represent and process, at scale, complex primary sources that vary widely in digitization granularity, refinement quality, and archival organization and curation practices. This paper presents the data representation framework designed for large-scale processing and indexing of historical newspapers and radio broadcasts developed within the Impresso project. Grounded in a structured characterization of the heterogeneity found in digitized historical media collections, it identifies the distinct dimensions along which collections diverge and the challenges they pose for a unified representation and processing framework. The framework navigates the competing demands of machine learning pipelines requiring uniform and lightweight document representations, information retrieval systems requiring well-defined indexable content units, user-facing interfaces requiring fidelity to original sources, and the need to return semantically enriched data to archival holders in interoperable formats. We describe the design principles guiding the framework and discuss how it reconciles these constraints across highly heterogeneous collections into a unified and research-ready corpus.