Overview
The FOSSIL (Footnote-based Open-access SSH Scientific Instance Labels) initiative addresses challenges in reference extraction from scholarly works in law and humanities. This project introduces an openly licensed, multilingual dataset comprising 96 annotated scholarly articles. This dataset contains over 7,600 footnote-embedded references. Complementing the dataset are a documented seven-annotator workflow, a collaborative web annotation tool named PDF-TEI Editor, and a specific specialization for Grobid tailored to footnote-based citations.
Research Context
Citation extraction tools commonly operate effectively with structured end-of-document bibliographies, which are prevalent in the natural sciences. However, scholarship within law and the humanities typically incorporates references primarily within footnotes. These footnote citations often interleave bibliographic data with commentary and cross-references. Furthermore, the presentation of bibliographic data in footnotes exhibits considerable variation across different languages and stylistic conventions. This context highlights a scarcity of suitable gold-standard resources designed for this specific format and domain.
Approach
The FOSSIL project developed a dedicated dataset and workflow to address the identified limitations. The dataset, FOSSIL, consists of 96 scholarly articles annotated for footnote-embedded references. The annotation process was supported by the PDF-TEI Editor, a collaborative web annotation tool. A seven-annotator workflow was documented to ensure consistency and quality in the annotation of these scholarly articles. This framework was then utilized to specialize Grobid, an existing citation extraction tool, for the specific characteristics of footnote-based citations found in law and humanities scholarship.
Findings
An end-to-end evaluation was conducted to assess the performance of the specialized pipeline. This evaluation compared the specialized Grobid pipeline against the default Grobid configuration. The results indicated that the specialized pipeline nearly doubled the extraction quality for footnote-based citations. Specifically, the micro-F1 score improved from 0.36 with default Grobid to 0.72 with the specialized pipeline. This substantial improvement in extraction quality was primarily attributed to an enhanced recall rate. The evaluation also indicated that considerable headroom for further improvement remains, particularly concerning cross-references and footnotes that combine various content types.
Why This Matters
The FOSSIL dataset and specialized Grobid pipeline provide a resource addressing the unique challenges of extracting citations from footnotes in law and humanities scholarship, where existing tools designed for scientific bibliographies are less effective. This effort contributes to the development of tools capable of handling the diverse and complex nature of references within these academic disciplines.