GROBID
A machine learning library that extracts, parses and restructures raw scholarly documents such as PDFs into structured XML/TEI-encoded data, focused on technical and scientific publications.
GROBID extracts and parses header bibliographic information, such as title, abstract, authors, affiliations and keywords, from a PDF article, and extracts and parses its references, scoring around .87 to .90 F1 against independent PubMed Central and bioRxiv test sets using its deep-learning citation model. It also recognizes and resolves citation contexts against the full bibliography, scoring between .76 and .91 F1 depending on the evaluation set.
It extracts and structures full text from PDF articles, including document segmentation and text-body structuring such as paragraphs, section titles, footnotes, figures, tables and data-availability statements; records PDF coordinates so bounding boxes can drive an 'augmented' interactive PDF; and parses references, names, and affiliation and address blocks in isolation, each scoring above roughly .90 F1 at the instance level.
It is written in Java, started as a hobby project in 2008 and has been open source since 2011 with continued support from Inria, and is licensed under Apache-2.0.