FootnotesExtractor: An Open-Source, Interactive, Multi-Engine System for Bibliographic Reference Extraction and Parsing from Scholarly Book Publications

We present an open-source system, Footnotes Extractor, for extracting bibliographic references and their components from scholarly book publications, with a focus on publications in the social sciences and humanities (SSH), a domain long underserved by existing citation-extraction tools. Our pipeline follows a multi-stage architecture, at each stage of which multiple interchangeable engines can be selected and combined: object detection localizes footnotes and reference regions on the page; OCR transcribes text from both scanned and digital-born documents; multipage consolidation reassembles footnotes and references spanning page or column boundaries; and NER-based models predict reference spans and their internal metadata components. This modular, multi-engine design allows us to select the best-performing combination for each family of documents and provides robust multilingual support, correctly handling the stylistic diversity of different publishers and citation conventions. A key feature is live, sequential editing: annotators can correct output at any stage, automatically triggering re-prediction in downstream stages and substantially reducing manual effort. Users can organize work into collections and groups for project management. The system resolves "ibidem" and other short-form references to their full citations, and enriches metadata via Google Books, Crossref, and OpenAIRE. Batch processing and a REST API support large-scale workflows, with export to annotated PDFs, JSON, or citation styles (BibTeX, APA, Chicago). In evaluations, Footnotes Extractor achieves Precision = 0.793, Recall = 0.867, and F1 = 0.815, outperforming established tools such as GROBID and CERMINE, particularly on SSH monographs. All code, trained models, and datasets are freely available at: https://github.com/FootnotesExtractor.