Over 16 million pages of historic US newspapers annotated with bounding boxes, predicted visual types, and OCR content. This dataset is a Parquet-converted version of the Newspaper Navigator dataset from the Library of Congress, originally released as JSON. The work was carried out as part of a project by Benjamin Germain Lee et al.
Use Cases
- Train OCR models based on the extracted text content mentioned in the description
- Analyze visual content distribution in historical newspapers based on predicted visual types (e.g., photographs, maps)
- Build multimodal search engines for historical archives based on combined OCR and image annotations
- Study layout and composition of historic newspaper pages based on bounding box annotations
Strengths
- Over 16 million annotated pages provide a large-scale corpus
- Data originates from the authoritative Library of Congress collection
- Includes multimodal annotations (OCR text and visual type predictions)
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
- Description metadata is limited; actual data quality requires manual inspection after download
Provenance
- Source
- Library of Congress
- Collection Method
- Project by Benjamin Germain Lee et al.; original data released as JSON, converted to Parquet.
- Freshness
- Last updated 2025-05-20 11:18:03; freshness should be verified
- Geography
- United States