Prosopography AI: The Extraction Pipeline

Prosopography AI is the use of automated tools — handwritten text recognition, named-entity recognition, coreference resolution, and probabilistic record linkage — to extract structured data about groups of historical people from archival records. This article sets out the pipeline stage by stage, names each stage's failure mode, and shows why the accuracy of everything downstream is bounded by the quality of the transcription it inherits. If you are building person data from manuscripts, the transcription layer is where the outcome is decided.

Prosopography AI is not a single model that reads an archive and returns a table of people. It is a pipeline of distinct stages, each with its own failure mode, and the accuracy of every downstream stage is bounded by the quality of the transcription it inherits. Used with human verification at the points that matter, AI can turn a bottleneck into a draft. Used without it, it manufactures plausible people who never existed.

That is the honest shape of the problem, and it is worth setting out carefully, because the field is currently full of claims that skip the middle.

What prosopography actually is

The term is Lawrence Stone's, from his 1971 essay in Daedalus: prosopography is the collective biographical investigation of a group of people through the common characteristics they exhibit in the records. The unit of analysis is the group, not the individual life. Where biography narrates one person, prosopography asks what a population of bailiffs, or ordinands, or correspondents held in common — offices, kinship, place, patronage — and reads those regularities as historical evidence.

Done at scale, this becomes a data problem. You are reading many records, identifying the people named in them, working out which mentions refer to the same person, and assembling those observations into a structure you can query. The classic large projects — the Prosopography of Anglo-Saxon England, People of Medieval Scotland, the Clergy of the Church of England Database, Six Degrees of Francis Bacon — are each an attempt to hold thousands of persons and their attested relations in one interrogable body. The China Biographical Database holds roughly 658,000 individuals. These are the ambitions AI is now being asked to accelerate.

The factoid model, and why a spreadsheet is not a prosopography

Before touching a single tool, it is worth internalising the data architecture the field converged on, because it dictates what "person data" should look like. The dominant model is the factoid model, developed by John Bradley and Michele Pasin at King's College London. It separates three entities:

Person — a scholarly inference: the historical actor you posit, assembled from many observations.
Source — the record that supplies an observation: a charter, a will, a tax roll, a parish register.
Factoid (or assertion) — the claim a source makes about a person at a point in time: held the office of bailiff of Reading, 1347.

The point of keeping these separate is that contradictory, weakly attested, and alternative observations can coexist as siblings rather than being flattened into a single "true fact." Two sources that disagree about a person's death date both survive in the model; the historian's inference sits at the person layer, traceable back through the factoids to the sources that ground it.

This is why a flat spreadsheet — one row per person, one cell per attribute — is not a prosopography. It collapses every observation into a single asserted value, hides provenance, and silently resolves contradictions the sources never resolved. If you are building person data with AI, the factoid architecture is the thing the pipeline should feed, and the reason source integrity matters from the very first stage.

The extraction pipeline, stage by stage

Turning a shelf of manuscripts into queryable person data is not one operation. It is five, and treating them as one is the commonest way the work goes wrong. Each stage has a characteristic failure mode, and errors compound downstream.

Transcription (HTR / OCR)

The image becomes text. This is the foundation, and its failure modes are quiet: character substitutions, silent normalisation of archaic spelling, dropped abbreviations. On clean printed Latin-script material, transcription error rates sit in the low single digits; on early-printed or damaged material they climb to 10–30% and beyond. Manuscript hands the model has seen before land in the 3–8% range; genuinely unfamiliar hands, 10–20% and up.

The reason this stage governs everything is not obvious until you have measured it. Studies injecting synthetic transcription noise find that degradation at 5–20% character error rate costs downstream named-entity recognition somewhere between 10 and 25 absolute F1 points on Latin-script historical documents. The variants that transcription destroys — the original spelling of a surname, the abbreviation mark on a place-name — are precisely the features NER and record linkage depend on. This is the single most important thing to understand about the whole enterprise: transcription is not separate from person data. It is the substrate. A fuller treatment of how to read the accuracy figures lives in the practical guide to handwritten text recognition.

Named-entity recognition (NER)

The model locates and classifies person mentions in the text. Failure modes: missed entities, boundary errors (grabbing half a name), and type confusion, all made worse by spelling variants and transcription noise.

It is tempting to assume NER works about as well on historical text as it does on modern newswire. It does not. The HIPE shared tasks — the standard benchmarks for named-entity work on historical newspapers — put best-system coarse person-detection F1 in the mid-0.80s to low-0.90s for French and German on clean text, dropping to the mid-0.70s on noisy-OCR English, with a further 10–15 point gap between clean and noisy variants. Those are meaningful scores, but they are 10–20 points below modern-newswire benchmarks, and entity linking lags mention detection by a substantial further margin. Historical NER is a distinct, harder problem; the field's answer is historical-language models plus budgeted human correction, not off-the-shelf systems. The mechanics of this stage are covered in depth in the guide to named-entity extraction from historical documents.

Coreference (within-document)

Once mentions are found, you cluster the ones that refer to the same person inside a single document — "the said John," "he," "the aforesaid bailiff." Failure mode: split entities (one person read as several) or merged entities (several read as one). Off-the-shelf neural coreference systems transfer poorly to historical formulae and dialogue; rule-augmented and in-corpus systems do better.

Entity resolution / record linkage (cross-document)

This is the hard centre of the discipline: deciding that the John Cooper in a 1347 tax roll and the John Cooper in a 1352 charter are — or are not — the same person. The foundational framework is Fellegi and Sunter's 1969 probabilistic theory of record linkage, which underlies most modern systems, typically with blocking to make the comparison tractable.

It is not a solved matching problem, and anyone selling it as one is misleading you. The Abramitzky–Boustan–Eriksson method for US historical censuses deliberately uses conservative thresholds to buy precision at the cost of recall: the 1940 full-count linkage reached roughly 80% coverage of native-born men and 60% of native-born women, with documented differential linkage by race, ethnicity, name commonness, and enumeration era. Two lessons follow. First, report precision and recall from a hand-labelled sample rather than trusting the tool's own confidence. Second, document who your method systematically misses — because it does, and the omission is rarely random. These frameworks were also built and validated on US census data; their behaviour on pre-modern European sources, with their different naming conventions and survival patterns, is far less established.

Structured modelling

Finally, the resolved persons, sources, and relations are encoded — in TEI XML using the `<listPerson>`, `<person>`, `<persName>`, and `<relation>` elements of Chapter 13; in a relational database; or in an RDF graph built on CIDOC-CRM's event-based ontology, which suits factoid-aware modelling. Failure mode: modelling choices that flatten uncertainty — a schema that forces one death date when the sources give two.

Where LLMs help, and where they fabricate people

The obvious temptation, in 2025, is to hand a page to a general-purpose chatbot and ask it for the people. This is worth addressing directly, because it is the first thing many researchers reach for.

Large language models are genuinely useful at this work in one role: draft extraction. Given clean text, a GPT-4-class model will propose a structured list of persons and relations with stronger recall than a rule-based system, which is a real head start on annotation. But the evaluations are unambiguous about the limits. Zero-shot GPT-3.5/4 reaches only around 0.55–0.65 strict F1 on the HIPE historical tasks — well below supervised baselines — and, more dangerously, it fabricates: plausible person names, dates, and locations that are simply not in the source, a tendency that worsens on Latin-language commentary and other material far from the model's training distribution.

This is the failure mode that should worry a prosopographer most, because it is invisible. A garbled transcription looks garbled; you catch it. A fabricated bailiff named for a plausible fourteenth-century Englishman, slotted into an otherwise sound extraction, looks exactly like data. It propagates into your person layer, your network graph, your published argument, and nothing about its surface flags it as invented. The broader pattern of fluent-but-wrong AI output is the defining risk of putting general models into a historical workflow.

The consensus practice, accordingly, treats LLM extraction as pre-annotation only: a draft a human verifies against the source, never data published raw. That is a workable division of labour — provided the text the model is reading is itself faithful to the page. Which returns us to the substrate.

Getting the substrate right: transcription that preserves what NER needs

Everything above rests on a claim that deserves stating plainly: the transcription stage is not preliminary housekeeping to be rushed through before the interesting work begins. It sets the ceiling on everything downstream. A 15-point NER penalty for a mediocre transcription is not recoverable by a better NER model; the information the model needs was destroyed before it arrived.

Two properties matter at this stage, and they pull in the same direction. The first is accuracy on the actual hands and founts you are working with — early-modern secretary and court hands, blackletter, the long s, the abbreviations that carry real linguistic content. The second, less discussed but just as consequential, is source integrity: a transcription that preserves original spelling, abbreviation marks, and scribal variants rather than silently modernising them. NER and record linkage key on exactly those variants. A tool that "helpfully" normalises Willm to William, or expands a contraction without marking it, has thrown away the evidence your linkage stage was going to use to distinguish two men with the same modernised name.

This is the stage where a purpose-built handwritten text recognition system earns its place in a prosopography pipeline, and it is the one stage of the five that I can speak to concretely. Leo's transcription model, ATR-1, is built for Latin-script material — any language written in the Roman alphabet, English and French and German and Dutch and Spanish and Italian and Latin among them, not the Latin language specifically — and it is trained to transcribe what is on the page rather than to smooth it into modern, plausible prose. The abbreviation stays an abbreviation; the archaic spelling survives; the strikethrough and the marginal addition are preserved rather than tidied away. That discipline is exactly what the factoid model and the linkage stage need from their substrate.

On accuracy, the one head-to-head figure worth citing: on a randomised 97-image sample of early-modern English manuscripts from the Folger Shakespeare Library, ATR-1 scored roughly 5% character error rate — 61% fewer errors than the next-best model tested (Transkribus's Text Titan I at ~13%, Claude Opus ~23.3%, Gemini 2.5 Pro ~24.8%, GPT-4.1 ~56.7%), at ATR-1's release. Given the 10–25 F1 points that hang on the transcription layer, the gap between a 5% and a 13% error rate is not a marginal convenience; it is the difference between a workable person-data substrate and one that quietly poisons the wells downstream. Leo also errs recoverably — on characters and words you check against the image shown beside the text — rather than with the fluent fabrications a general model produces. The transcriptions then export to TEI XML, which is where a prosopography project's structured modelling begins.

What Leo does not do is the rest of the pipeline. It transcribes and organises; it does not resolve entities across documents or build your network graph. That is honest, and it is the right boundary: the transcription stage is where a specialist model changes the outcome, and the linkage stage is where your judgment and the established probabilistic methods do.

A working sequence

Putting the pieces together, the current best-practice pipeline for building person data from archival records looks like this:

Capture and transcribe the records with a specialist HTR model, preserving original orthography and abbreviations, and treating transcription quality as a first-class evaluation variable — measured, not assumed.
Extract entities with a historical-language NER model, or an LLM as a draft generator, budgeting for human correction rather than trusting the output.
Resolve coreference within documents, preferring rule-augmented or in-corpus systems over off-the-shelf neural ones.
Link records across documents with a probabilistic method, conservative thresholds, and a hand-labelled sample to estimate precision and recall — and a plain statement of who the method misses.
Model the resolved persons, sources, and assertions in a factoid-aware structure — TEI, a relational schema, or an RDF graph — that preserves uncertainty rather than flattening it, and analyse it in a purpose-built tool such as Nodegoat or Palladio.

At every join between these stages, a human verifies. That is not a concession to immature technology; it is the structure of the discipline. Prosopographic identity is a scholarly inference, contestable and historiographic, and no pipeline resolves that — it only supplies the evidence faster.

The gain AI offers is real and worth having: the transcription bottleneck that once gated the whole enterprise can now be measured in weeks rather than years, and the drudgery of first-pass entity extraction can be handed to a draft generator. But the value of a prosopography still lives where it always did — in the quality of the inferences you draw and the defensibility of the sources beneath them. The tools change the speed of the reading. They do not change what it means to have read carefully.

Frequently Asked Questions

What is prosopography AI?

Prosopography AI is the use of automated tools — handwritten text recognition, named-entity recognition, coreference resolution, and probabilistic record linkage — to extract structured data about groups of historical people from archival records. It is not a single model that reads an archive and returns a table of people. It is a pipeline of distinct stages, each with its own failure mode, and the accuracy of every downstream stage is bounded by the quality of the transcription it inherits. Used with human verification at the points that matter, it turns a bottleneck into a draft; used without it, it manufactures plausible people who never existed.

What is the factoid model in prosopography?

The factoid model, developed by John Bradley and Michele Pasin at King's College London, is a data architecture that separates three entities: the person (a scholarly inference assembled from many observations), the source (the record supplying an observation, such as a charter or parish register), and the factoid (the claim a source makes about a person at a point in time). Keeping these separate lets contradictory and weakly attested observations coexist rather than being flattened into a single asserted value. This is why a flat spreadsheet is not a prosopography: it hides provenance and silently resolves contradictions the sources never resolved.

Why does transcription accuracy matter so much for extracting person data?

Transcription accuracy sets the ceiling on everything downstream, because the whole pipeline inherits whatever the transcription produces. Studies injecting synthetic transcription noise find that degradation at 5–20% character error rate costs downstream named-entity recognition somewhere between 10 and 25 absolute F1 points on Latin-script historical documents. The variants transcription destroys — the original spelling of a surname, an abbreviation mark on a place-name — are precisely the features named-entity recognition and record linkage depend on. A 15-point penalty for a mediocre transcription is not recoverable by a better later model; the information was destroyed before it arrived.

Can I just ask ChatGPT to extract the people from a historical document?

You can use a large language model for draft extraction, but not for finished data. Given clean text, a GPT-4-class model proposes a structured list of persons and relations with stronger recall than a rule-based system, which is a real head start on annotation. But zero-shot GPT-3.5/4 reaches only around 0.55–0.65 strict F1 on the HIPE historical tasks, and more dangerously it fabricates plausible names, dates, and locations not in the source. That failure mode is invisible: a fabricated bailiff looks exactly like data. Treat LLM output as pre-annotation a human verifies against the source, never data published raw.

What is record linkage in prosopography, and is it a solved problem?

Record linkage is the process of deciding whether a person named in one document — say, a John Cooper in a 1347 tax roll — is the same person named in another. It is the hard centre of the discipline and it is not solved. The foundational framework is Fellegi and Sunter's 1969 probabilistic theory, usually paired with blocking to make comparison tractable. Conservative methods buy precision at the cost of recall and systematically miss people — one census linkage reached roughly 80% coverage of native-born men and 60% of native-born women. Report precision and recall from a hand-labelled sample, and document who your method misses.