Matteo Cargnelutti, Catherine Brobston, Eben English, Jake Sadow, Kacie Bailey, Greg Leppert, Amanda Watson, Jessica Chapel, Jonathan Zittrain
6 min
Historical newspapers provide a rich record of public life, but their dense, irregular, and noisy layouts make computational access challenging. Traditional OCR engines often fail on complex newspaper columns and decorative headings, while processing full scans through vision-language models degrades transcription accuracy. To address this, the researchers developed the Institutional Newspapers Pipeline, a modular system designed in collaboration with the Boston Public Library to extract high-quality, structured datasets from historical newspaper scans while remaining computationally frugal enough to run on workstation-level hardware.
The pipeline processes each scan through a multi-step workflow. First, a YOLO-based object detection model segments newspaper scans into type-agnostic rectangular crops—bounding boxes around uninterrupted flows of text or visual elements. Separating segmentation from downstream classification allows the system to utilize limited training annotations effectively without class imbalance issues. Subsequent steps perform optical character recognition on each crop using specialized vision-language models, followed by text analysis, crop-type classification, reading order detection, named entity recognition, subject classification, language detection, and pre-computed embeddings generation.
Running this pipeline against a portion of the Boston Public Library's holdings between 1795 and 1930 yielded 83.1 million individual crops across 1,473,635 public domain newspaper scans. The resulting OCR output amounts to 16.3 billion tokens. Evaluation of the segmentation model demonstrated high precision and recall, with Eigen-CAM heatmaps confirming that the model learns structural cues such as layout boundaries and column structures rather than incidental scan features. Mid-19th-century scans exhibited the highest density in terms of average crops per scan.
This work provides an open-source, production-ready pipeline and dataset that overcomes the historical inaccessibility of digitized newspapers. By producing atomic, high-quality representations of historical newsprint, the pipeline improves patron access and supplies clean data for digital humanities research and the pre-training of large language models on historical text.
Historical newspapers are an abundant record of public life, but their dense, irregular and sometimes noisy layouts make computational access to these materials both challenging and limited. We present the Institutional Newspapers Pipeline, a modular system we jointly designed with Boston Public Library to extract high-quality, structured datasets from historical newspaper scans. It was architected so that each step remains interpretable and customizable, and so that the pipeline as a whole remains computationally frugal enough to run on workstation-level hardware. The pipeline runs each scan through a multi-step process: it segments scans into individual type-agnostic crops and performs OCR on each resulting segment before then performing text analysis, type classification, reading order detection, named entities recognition, subject classification, language detection, and pre-computed embeddings generation on every crop. We ran this pipeline against a portion of Boston Public Library's holdings and released the results as an open dataset. The optical character recognition (OCR) output represents 16.3 billion o200k_base tokens across 83.1 million individual crops, extracted from 1,473,635 public domain newspaper scans published between 1795 and 1930. This report describes our methods for each processing step, the small models we trained, as well as the evaluation results and dataset-scale measurements we collected in the process. It accompanies the release of the pipeline, models, and dataset. We position this work as a substantial step towards unlocking high-quality data from tens of millions of newspaper scans.
Alex: And the compute cost? Running a three-billion-parameter model on every crop from over a million scans sounds like it adds up fast.
Sam: It does, and that's where the 'frugal' design philosophy becomes load-bearing. They use aggressive batching and inference caching through vLLM, which substantially reduces redundant computation. The claim is that the full pipeline runs on workstation hardware — no cluster required. The output is around sixteen billion tokens of structured text extracted from over one-and-a-half million page scans. Whether that throughput figure holds across different hardware configurations isn't something the paper stress-tests, but the architecture is clearly designed with that constraint in mind from the start.
Alex: So the contribution isn't really any single model — it's the modular design that makes the whole thing tractable.
Sam: That's the right read. Segmentation, OCR routing, and classification are kept as distinct, composable steps. That modularity matters for two reasons. First, it enables parallelization — crops from different pages can be processed concurrently without any dependencies between them. Second, it makes the pipeline interpretable and maintainable. If the segmentation model improves, you swap it in without touching the transcription layer. If a better VLM comes along, same story.
Alex: Where would a careful reviewer push back?
Sam: A few places. The paper doesn't include a systematic evaluation of OCR quality against a held-out ground truth — there's no character error rate or word error rate reported against a labeled benchmark. The output scale is impressive, but without that, it's hard to know how much of those sixteen billion tokens are clean versus noisy. There's also the question of generalizability: the pipeline is tuned for American historical newspapers from a specific era and collection. How it performs on, say, non-English archives or different print technologies is an open question the authors don't address.
Alex: So the headline result — the scale of the output — is real, but the quality characterization is underdeveloped.
Sam: That's a fair summary. The architectural choices are well-motivated and the compute efficiency argument is credible. But for researchers who want to build on this corpus, the absence of a rigorous quality benchmark is a meaningful gap. You'd want to run your own spot-checks before treating the extracted text as a reliable input to downstream analysis.
Alex: It's a useful infrastructure contribution, even if the evaluation leaves room for follow-up work.
Sam: Exactly. And for computational historians or digital humanities researchers, the practical value is real — this kind of pipeline lowers the barrier to working with archival material that was previously locked behind unusable scans. The modular design also means the community can iterate on individual components rather than treating the whole system as a black box.
Alex: Thanks for walking through it. If you're working with historical corpora or building OCR pipelines for archival data, this paper is worth a close read — both for what the architecture gets right and for the evaluation questions it leaves open. Thanks for listening to ResearchPod.