PubMed-OCR: PMC Open Access OCR Annotations

TL;DR

PubMed-OCR addresses the need for faithful text-layout grounding in scientific literature by providing an OCR-first corpus derived from the PubMed Central Open Access subset and annotated directly from page images with word-, line-, and paragraph-level bounding boxes. It processes 209.5K documents using Google Vision OCR on page images rendered at 150 DPI, producing standardized JSON outputs and line-reconstructed layouts that avoid PDF/XML alignment errors. The dataset comprises about 1.5 million pages and roughly 1.3 billion words, enabling layout-aware modeling, coordinate-grounded QA, and robust evaluation of OCR-dependent pipelines, with rich statistics on journals and layout features. While offering broad utility and reproducibility, the work notes limitations such as dependence on a single OCR engine, heuristic line reconstruction, and lack of character-level or mathematical structure annotations, and it invites community extensions and audits to improve coverage and accuracy.

Abstract

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.

Figures (8)

• Figure 1: Distribution of number of words (left), lines (middle), and paragraphs (right) per page. $\mu$ indicates the mean, $M$ indicates the median, and $\sigma$ is the standard deviation. Each distribution is truncated at or below the 99.5th percentile to visualize the core probability mass instead of the long tail.
• Figure 2: Distribution of number of words (left), lines (middle), and paragraphs (right) per document. $\mu$ indicates the mean, $M$ indicates the median, and $\sigma$ is the standard deviation. Each distribution is truncated at or below the 99th percentile to visualize the core probability mass instead of the long tail.
• Figure 3: Top 20 journals represented in PubMed-OCR. The top 3 journals account for $\sim$23% of all documents included.
• Figure 4: Two example pages from PubMed-OCR, overlaid with layout detection classes predicted by PP-DocLayout. On the left, we have a page with a seal alongside high-density text (with formulas embedded within the text). On the right, we have a page with many tabular outputs, code snippets, and other text.
• Figure 5: A sample page from PubMed-OCR exhibiting a variety of features: aside text, charts, captions, and formulas.