Source-linked AI summary
PubMed-OCR: PMC Open Access OCR Annotations
Hunter Heidenreich, Yosheb Getachew, Olivia Dinica, Ben Elliott
TL;DR
Scientific-document research needs open, grounded OCR resources that preserve both text and page layout, including for scanned pages. PubMed-OCR constructs such a corpus from PMCOA page images using Google Vision OCR and releases structured annotations at multiple granularities. It contains 209.5K documents and 1.5M pages, while its coverage and annotations have documented scope limitations.
Problem
Scientific articles require grounded OCR because dense mathematics, references, tables, and figures challenge document understanding, while faithful OCR preserves spatial provenance for downstream systems.
Method
PubMed-OCR samples PMCOA Open Access PDFs, renders pages as images, applies Google Vision OCR, and extracts word, line, and paragraph bounding boxes into released JSON annotations.
Results
209.5K documents and 1.5M pages comprise the release, with 39.5 paragraphs, 106.3 lines, and 844 words per page on average.
Takeaways & Limitations
The corpus supports layout-aware modeling, coordinate-grounded question answering, attributed generation, training, and evaluation on heterogeneous scientific articles.
Takeaways & Limitations
The corpus relies on one OCR engine, reconstructs lines heuristically, omits character boxes and explicit mathematical or figure/table structure, and reflects PMCOA licensing and journal distribution.
Abstract
from arXiv · showhide
PubMed-OCR is an OCR-centric corpus of scientific articles derived from PubMed Central Open Access PDFs. Each page image is annotated with Google Cloud Vision and released in a compact JSON schema with word-, line-, and paragraph-level bounding boxes. The corpus spans 209.5K articles (1.5M pages; ~1.3B words) and supports layout-aware modeling, coordinate-grounded QA, and evaluation of OCR-dependent pipelines. We analyze corpus characteristics (e.g., journal coverage and detected layout features) and discuss limitations, including reliance on a single OCR engine and heuristic line reconstruction. We release the data and schema to facilitate downstream research and invite extensions.
1 Introduction
Real-world document understanding depends on robust OCR and open data, especially for scientific literature. PubMed-OCR addresses this need by directly annotating page images with structured OCR supervision.
- Open corpora improve models, reproducibility, participation, and benchmarking across language, multimodal, and document-processing research.
- OCR translates optical signals into discrete text, making document information accessible to text-only models.
- PubMed-OCR annotates page images with Google Vision OCR, producing word-, line-, and paragraph-level supervision.
- The corpus bypasses alignment between digital PDFs and JATS XML, avoiding parser noise, heuristic dependencies, and missed text from scanned documents.
- The released resource supports model development, benchmark curation, and related research on scientific documents.
2 Related Work
Prior resources provide large-scale document, layout, or table annotations, but many rely on parser-derived text or lack grounded OCR. PubMed-OCR instead derives structured supervision directly from page images, including scanned scientific pages.
- GROTOAP2 provides hierarchical XML annotations across zones, lines, words, and characters, with 22 zone labels and two-point bounding boxes.
- PubLayNet aligns PMCOA XML with PDFMiner output to produce ∼3.5M region annotations across ∼360K pages and five canonical classes.
- PubTables-1M annotates 948K tables across 575K pages at table, row, column, and cell levels, with PDF and image coordinates.
- PubMed-OCR bypasses PMCOA XML and derives word-, line-, and paragraph-level supervision directly from page images, including scanned pages without text overlays.
- Scientific articles add dense mathematics, fine-grained references, and extensive table and figure usage to general document-processing challenges.
- Grounded OCR preserves spatial provenance for layout-aware modeling and verifiable attribution, unlike plain text recognition that serializes pages into one sequence.
- Grounded OCR can be deterministically linearized when needed, while retaining more flexible word-, line-, and paragraph-level structure.
3 PubMed-OCR Dataset
PubMed-OCR constructs an OCR-first corpus by annotating rendered PMCOA page images and releasing grounded word-, line-, and paragraph-level artifacts. Its scale, granularity, journal breadth, and layout diversity support document-understanding research while retaining important coverage and annotation limitations.
- OCR processing: Google Vision OCR processes 150-DPI page images, extracting word polygons and deriving paragraph boxes from their vertices.The corpus emits page-level JSON containing words, lines, paragraphs, and image metadata.
- OCR processing: Heuristic line reconstruction groups vertically aligned words, splits groups by paragraph, and orders words left-to-right.Words join groups when their minimum and maximum y-coordinates differ by no more than 5 pixels.
- Corpus scale: 209.5K documents and 1.5M pages comprise the release, averaging 7.4 pages per document.
- Corpus statistics: 844 words, 106.3 lines, and 39.5 paragraphs occur per page on average, while documents average 6,229.6 words, 784.9 lines, and 291.3 paragraphs.
- Corpus statistics: Despite fewer documents and pages than OCR-IDL, PubMed-OCR has almost 4x as many line annotations and 10x as many word annotations.The comparison caption warns that OCR-IDL's published totals imply an order of magnitude more words and lines than its manuscript reports.
- Journal distribution: The dataset represents 2,478 journals, although its top three titles account for roughly 23% of documents.Singleton journals comprise roughly 25.7% of journals but only 0.3% of documents.
- Layout diversity: Layout detection on 40,000 sampled pages finds formulas on ∼25%, images on ∼22%, charts on ∼16%, and tables on ∼18%.Figure 4 illustrates pages combining dense text, formulas, tabular data, algorithmic definitions, titles, and other features.
4 Conclusion
PubMed-OCR provides an OCR-first, openly accessible corpus with grounded layout annotations and corpus statistics, while retaining important annotation and coverage limitations.
- Conclusion: PubMed-OCR exposes paragraph-, line-, and word-level bounding boxes directly from page images in a compact, standardized format.The resource is derived from the PubMed Central Open Access subset and supports training and evaluation.
- Conclusion: The corpus supports layout-aware modeling, grounded question answering, and attributed generation on scientific literature.Its corpus-level distributions characterize scale and diversity across journals and article lengths.
- Limitations: PubMed-OCR relies on a single OCR engine and heuristically reconstructed line annotations, which may bias reading order and grouping.It also lacks character-level boxes and explicit mathematical, figure, or table structure representations.
- Limitations: The corpus coverage reflects PMCOA’s license and journal distribution, requiring these constraints to be considered in reporting and experimental splits.The authors explicitly connect these boundaries to dataset use and evaluation design.
- Conclusion: The top 3 journals account for ∼23% of all documents included.This indicates a concentrated journal distribution within the corpus.
A More Examples
Figures 5–8 provide additional PP-DocLayout samples illustrating varied page layouts and structured annotations, while also exposing a table misidentification.
- Figures 5–8 show additional PubMed-OCR pages with layout detection annotations from PP-DocLayout.
- Figure 5 illustrates a page containing aside text, charts, captions, and formulas.
- Figure 6 presents a complex scientific table in which a second tabular section is misidentified as an algorithm.
- Figure 7 shows dense text, a table of contents, and structured detection of paragraph and document titles.
- Figure 8 depicts an image-dense page with structured captions describing details and intended interpretation.