Layout-aware text extraction from full-text PDF of scientific articles

Area of Science:

• Computational Biology
• Bioinformatics
• Scientific Publishing

Background:

• Portable Document Format (PDF) is the standard for online scientific publications.
• Extracting text from PDFs in a layout-aware manner is challenging for text mining and biocuration systems.
• Existing methods struggle to accurately process the structural information within PDF scientific articles.

Purpose of the Study:

• To introduce the Layout-Aware PDF Text Extraction (LA-PDFText) system.
• To facilitate accurate, layout-aware text extraction from PDF research articles for text mining applications.
• To provide a baseline system for future research in multi-modal content extraction.

Main Methods:

• The LA-PDFText system employs a three-stage process: detecting contiguous text blocks using spatial layout, classifying blocks into rhetorical categories via a rule-based method, and stitching blocks in the correct order.
• Spatial layout processing is used to identify and group text blocks.
• A rule-based classifier categorizes text blocks into logical units such as title, abstract, and body text.

Main Results:

• The LA-PDFText system achieved high performance in identifying text blocks and classifying them into rhetorical categories, with Precision=0.96, Recall=0.89, and F1=0.91.
• Evaluated the accuracy of the block detection algorithm and compared text extraction accuracy against the PDF2Text system and PubMed Central.
• Demonstrated the system's effectiveness in extracting text from section-wise grouped blocks.

Conclusions:

• LA-PDFText is an open-source tool that enables accurate text extraction from full-text scientific articles.
• The system provides a foundational solution for enhancing biomedical text mining and biocuration.
• The source code is publicly available for further development and application.