Marebito: citation matching interface for the digital library zbMATH open

Abstract zbMATH Open is a leading curated bibliographic database covering approximately five million publications in mathematics. This paper presents Marebito, a production citation-matching system developed for zbMATH Open, and documents its architecture, domain-adapted feature design, and operational performance at scale. Marebito combines bibliographic parsing with GROBID, candidate retrieval with Elasticsearch, and a supervised Random Forest trained on domain-specific string-similarity features. Evaluation on internal zbMATH Open datasets yields precision and recall exceeding 98% on modern DOI-indexed records. Evaluation across two bibliographically distinct subsets reveals that classifier performance is strongly conditioned on metadata completeness, particularly title and author string quality, rather than DOI availability alone, documenting a concrete challenge for citation matching in mathematical literature. A preliminary comparison with Crossref on a manually verified subset of 2,037 entries illustrates the limitations of general-purpose matchers on specialist mathematical bibliography. A neural reranking baseline confirms that Elasticsearch-derived hard negatives are essential for abstention. We also discuss downstream bibliometric applications within zbMATH Open.