Citation extraction

Finding and structuring the citations — footnote markers, bibliographic entries, in-line references — within a document so they become queryable objects rather than prose.

Why it matters for your research. Once citations are extracted, the corpus becomes a graph: works that cite each other, authors who cluster by whom they cite, gaps in a citation chain that signal a missing source. For academic-historical corpora this is one of the highest-value analyses available.

In Archēglyph. Under Later on the project roadmap; non-trivial because citation formats vary wildly across time and field.

Not to be confused with. NER finds people and places; citation extraction finds the references — often overlapping entities, but structured as pointers into the citation graph.

Why it’s harder than it looks. A citation marker in a 19th-century periodical might be a superscript number, a bracketed year-author pair, an unmarked footnote keyed only by its position on the page, or a marginal note added by a later reader. Modern citation parsers like GROBID are trained mostly on contemporary journal PDFs with consistent reference-list formatting; historical and archival material breaks those assumptions in ways that need per-corpus tuning rather than a single model that works everywhere.

A worked example. Take a footnote reading Cf. Smith, op. cit., p. 42. Structuring it well means resolving op. cit. back to the earlier full citation it refers to, not just tagging “Smith” as a person and “42” as a number. Get that resolution right across a whole volume and you can answer “which sources does this author return to most” without a human tracing every footnote by hand — get it wrong and the citation graph silently drops every ibid. and op. cit. reference, which in some 19th-century scholarship is the majority of citations after the first chapter.

What “structured” buys you over plain search. A full-text search for an author’s surname finds every place the name is mentioned, including in the body prose, not only in citations. Structured citation extraction distinguishes a citing mention from a discussing mention — “Smith argues that…” versus a bare footnote reference — which matters if the research question is specifically about the citation network rather than every occurrence of a name. It also normalises variant renderings of the same source (a full citation on first use, a short form afterwards) to one canonical work, so counting “how many times is this source cited” doesn’t undercount because half the citations used the abbreviated form.

Relationship to the rest of the pipeline. Citation extraction sits downstream of layout analysis, which has to correctly identify footnote regions before there is anything to extract citations from, and downstream of sentence segmentation, since a citation marker’s boundaries depend on knowing where the sentence containing it ends.

What it would add to a dataset note. Once landed, citation extraction would let a dataset note report not just what a corpus contains but how densely it cites, and which works it returns to most — a different kind of summary from the topic or entity facets other plugins produce, closer to a bibliometric profile of the corpus than a description of its subject matter.

Why format variation is the hard part, not extraction itself. Recognising that a numeral or bracket is probably a citation marker is comparatively easy; resolving it to the correct entry in a bibliography formatted completely differently by a different press in a different decade is not. A tool that handles modern journal PDFs well can still fail on a 19th-century periodical whose editor invented its own shorthand, which is why this is scoped as an analysis plugin rather than a single universal parser.

What “queryable” means in practice. Once a citation is structured rather than left as prose, a researcher could ask a dataset directly for every citation to a given work, or filter a corpus down to documents that cite a particular source at all — questions a plain full-text search answers only approximately, because it cannot distinguish a citing mention from every other mention of the same name or title.

← Back to the glossary