# American Stories: Large-Scale Structured Text from Historical U.S. Newspapers > Research note on American Stories, a dataset of 1.14 billion content regions extracted from nearly 20 million historical U.S. newspaper scans via a modular deep learning pipeline. - Authors: Hunter Heidenreich - Published: 2026-03-15 - Paper: Preprint (arXiv) - Dataset: American Stories Dataset - Code: American Stories Pipeline - Source: Bevaya Labs, --- ## TL;DR American Stories applies a modular, cost-efficient deep learning pipeline (YOLOv8 layout detection, MobileNetV3 legibility classification, EfficientOCR) to nearly 20 million Library of Congress newspaper scans, producing 1.14 billion content region bounding boxes and 65.6 billion tokens of structured article text. The dataset is public domain content released under CC-BY-4.0. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$. The headline contribution is the American Stories dataset itself: 1.14B layout annotations and structured OCR text extracted from the Chronicling America collection. The paper's applications section, evaluation, and limitations all center on the dataset's fitness for downstream use. **Secondary:** $\Psi\_{\text{Method}}$. The paper develops a full extraction pipeline (layout detection, legibility classification, custom OCR, content association) and makes specific architectural choices (mobile-phone-class models) to meet a $60K cloud compute budget constraint. The pipeline is modular and released for reuse. ## What is the motivation? Library of Congress's Chronicling America project provides approximately 20 million historical newspaper scans with page-level OCR, but the digitized text does not respect page layout. Articles, headlines, captions, advertisements, and other regions are scrambled together at the page level. A non-trivial share of scans are illegible, introducing noise. These limitations prevent effective use of modern NLP methods, language model pre-training on historical English, and structured social science analyses that require article-level text. Existing open-source alternatives (notably Newspaper Navigator) detect visual content classes (photographs, illustrations, maps, comics, cartoons, ads, headlines) but do not detect article bounding boxes, which is the prerequisite for legibility filtering, custom OCR, and article association. ## What is the novelty? The core contribution is the end-to-end pipeline for converting newspaper scans into structured article text at scale: 1. **Layout detection:** YOLOv8-Medium trained on 2,202 labeled pages (48,874 layout objects) with 9 content classes: Article, Headline, Byline, Caption, Image, Advertisement, Table, Masthead, and Header. Achieves 91.3 mAP@50:95 on articles and 88.6 on headlines. 2. **Legibility classification:** MobileNetV3-Small classifies each text region as legible, borderline, or illegible. Trained on 979 labeled crops with weighted cross-entropy. No legible texts are misclassified as illegible in evaluation. 3. **OCR (EfficientOCR):** A character/word-level image retrieval approach using MobileNetV3-Small encoders trained with supervised contrastive loss. Words are recognized by nearest-neighbor lookup against an offline embedding index of 22,230 dictionary terms rendered across 43 fonts. When cosine similarity falls below 0.82, the system falls back to character-level recognition. This design meets the $60K compute budget while achieving 4.3% CER. 4. **Content association:** Rule-based methods associate headlines, bylines, and article bounding boxes using spatial overlap heuristics. Achieves 97.0 F1 on headline-article association. The pipeline prioritizes mobile-class architectures (YOLOv8, MobileNetV3) throughout, making it over an order of magnitude cheaper to deploy than alternatives like TrOCR. ## What experiments were performed? Evaluation uses four hand-constructed datasets: - **Full page scans** (10 pages, 597 content regions, 196,655 characters): end-to-end pipeline evaluation including content association (214 headline-article pairs). - **Day-per-decade sample** (50 lines per decade, 1850-1920): OCR evaluation across diverse historical periods. - **Random lines from Carlson et al.** (64 textlines): comparison with other detection frameworks and OCR engines. - **Legibility sample** (100 bounding boxes): legibility classifier evaluation. Key metrics: - End-to-end CER: 0.051 (0.044 with spellchecking) - OCR-only CER: 0.043 - Layout/line detection CER contribution: 0.012 - Content association F1: 97.0 - Legibility: 0/81 legible texts misclassified as illegible; 1/16 illegible texts misclassified as legible Downstream application comparisons against Chronicling America page-level OCR: - **Topic classification** (politics): RoBERTa on American Stories articles achieves F1 83.6 (article-level) and 96.0 (page-level), compared to 83.3 for neural methods on Chronicling America pages. - **Reproduced content detection**: Neural method on American Stories achieves ARI 86.2 (page-level), compared to 74.6 for Viral Texts on Chronicling America. ## What are the outcomes/conclusions? The American Stories dataset contains 1.14 billion content region bounding boxes across nearly 20 million scans, with 65.6 billion tokens of OCR text. Coverage spans all U.S. states, with content dating from the 17th century through the mid-20th century (bulk from early 1900s). The structured article texts enable analyses that are impossible with page-level OCR: article-level topic classification, reproduced content detection, and news story clustering. The authors demonstrate that neural methods on structured articles substantially outperform both neural and sparse methods on unstructured page text. Limitations the authors acknowledge: - Foreign-language newspapers are excluded (off-the-shelf OCR performs poorly on diverse scripts) - Only 3.8% of articles span multiple bounding boxes; these are not associated in the current release - Historical language reflects cultural biases of the period; the authors recommend against using the dataset for training generative models - The pipeline is optimized for cost over accuracy; swapping in larger models (ViT, two-stage detectors) could improve quality ## Reproducibility ### Models - **Layout detection:** YOLOv8-Medium, initialized from official pretrained checkpoint. Trained 100 epochs on 2,202 pages. Settings: `imgsz=1280, iou=0.2, max_det=500, conf=0.1`. - **Legibility classification:** MobileNetV3-Small (timm `mobilenetv3_small_050`). Trained 50 epochs on 979 crops. Resolution 256, lr 2e-3 with 0.1 decay every 20 epochs. Weighted CE loss [2.0, 1.0, 1.0]. - **Line detection:** YOLOv8-Small. 100 epochs on 4,000 synthetic lines, then 50 epochs on 373 annotated crops. Resolution 640. - **Word/character localization:** YOLOv8-Small. 100 epochs on 8,000 synthetic textlines, then 100 epochs on 684 annotated lines. Resolution 640. - **Word recognition:** MobileNetV3-Small (timm `mobilenetv3_small_050`). SupCon loss, $\tau=0.1$, AdamW, cosine annealing with warm restarts (max lr 2e-3). 50 epochs without hard negatives, 40 epochs with offline hard negative mining. Batch size 1024 ($m=4$ views $\times$ 256 words). - **Character recognition:** EffOCR-C (Small) from Carlson et al. - All model checkpoints available via Dropbox link in the GitHub repo. ### Data - **Source:** Library of Congress Chronicling America collection (~20M scans). Public domain (pre-1925 content). - **Layout annotations:** 2,202 pages labeled by undergraduate research assistants. Active learning sampling. 13% from Chronicling America, remainder from other public domain/off-copyright newspapers. - **OCR training data:** 43 digital fonts for synthetic rendering; silver-quality labels from EffOCR-C on random sample; small set of hand-labeled word crops. Dictionary of 22,230 terms. - **Output dataset:** 1.14B bounding boxes, 65.6B tokens. JSON format on HuggingFace Hub. CC-BY-4.0. ### Evaluation - CER (character error rate) via Levenshtein distance. Non-word rate as proxy metric on larger samples. - mAP@50:95 for layout and line detection. - F1 for content association. ARI (adjusted rand index) for reproduced content clustering. - Evaluation sets are small (10 full pages, 64-225 textlines). No error bars or multi-run analysis reported. ### Hardware - **Training:** Single NVIDIA A6000 GPU. - **Inference (pipeline deployment):** Azure F-series CPU nodes. Total compute budget: $60,000 USD for processing ~20M scans. - **Cost comparison:** EfficientOCR is reported as over an order of magnitude cheaper than TrOCR Base on cloud CPUs. ## BibTeX ``` ```bibtex @inproceedings{dell2023american, title={American Stories: A Large-Scale Structured Text Dataset of Historical {U.S.} Newspapers}, author={Dell, Melissa and Carlson, Jacob and Bryan, Tom and Silcock, Emily and Arora, Abhishek and Shen, Zejiang and D'Amico-Wong, Luca and Le, Quan and Querubin, Pablo and Heldring, Leander}, booktitle={Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track}, year={2023} } ``` ```