Human Capital in the Nordic Countries

HCNC · 90,028 lives, read out of 112 years of Norwegian student yearbooks

HCNC asks how the Nordic countries built their human capital: who received an education, what they did with it, and how advantage travelled between generations. Those are questions about individuals, so they need sources that record individuals — and the ones that do are almost all still on paper.

The project's first corpus, and so far its only one, is Norwegian. The Studentene fra… books — "The Students from…" — are a remarkable Norwegian publishing tradition. From the mid-nineteenth century, each volume documented an entire graduating class, usually published at the twenty-five-year reunion: every graduate's parentage, birthplace, education, career, marriages, children and travels, written down and printed. Response rates were above 90 per cent.

The result is one of the richest prosopographical sources anywhere — a century of the Norwegian educated elite, life by life. It is also 43,000 pages of dense print in shifting typefaces, which is why nobody had ever used it at scale.

What was built

A pipeline that turns those pages into a structured, queryable dataset: 90,028 records, each with thirteen fields covering birth, parentage, education, career, marriage, children and stays abroad. Three independent open vision–language models read every page; a language model then extracts structured fields from the text; records are aligned across engines, consolidated, and remaining disagreements adjudicated by an LLM judge with its reasoning logged.

Validation is what makes it usable. The dataset was checked against 37,712 biographies that had been transcribed independently by hand: names agree 99.3% of the time, full birth dates 97.2%, parents' names 98.7–99.1%. The weakest fields are birthplace (92.9%) and father's occupation (93.7%), where the difficulty is less OCR than history — "Kristiania" against a parish inside it, and a source culture of extreme abbreviation.

A finding worth the whole exercise

The project set out to use agreement between three OCR engines to correct errors. Measured against the hand-transcribed ground truth, that consensus made the data worse — it lost to simply keeping the best single engine's output on six of seven fields.

The diagnosis generalises. Majority voting assumes that the voters make independent errors. Two of the three engines descended from the same vision–language model family, so they failed in the same places at the same time and outvoted the more accurate third. Where all three agreed, father's occupation was wrong 12.5% of the time; where they disagreed and the majority decided, the chosen value was wrong 57% of the time.

Worse than the noise was the identity damage. Fuzzy merging of names across engines occasionally fused two adjacent people into one, attaching one graduate's biography to another's name — 939 records. For prosopography that is the error type that matters: field-level noise averages out, but a wrong identity contaminates every field in the record and every family link built on it.

The redesign takes the strongest engine as the backbone and uses the others only where their errors are demonstrably independent — full birth dates, where digit-level OCR mistakes really are decorrelated, and where consensus beats every single engine. Everywhere else, cross-engine agreement is retained not as a vote but as per-field confidence metadata, so researchers can filter to the level of corroboration their question requires.

What the data show

The descriptive results already trace a century of social change:

• Who got in. Farmers, merchants and teachers dominate among the fathers — the university working simultaneously as a machine for social reproduction and as a route out of the countryside.
• Women. The full arc of their entry into Norwegian higher education, from near-total absence before 1882 to more than a quarter of graduates by the 1940s. The early female graduates were an exceptionally privileged group: wealthier, more urban and more internationally mobile than the men. The most common recorded title for a woman graduate remained husmor — housewife.
• Looking outward. 36.8% of graduates recorded a stay abroad, above 45% for the classes of the 1890s and 1900s, with Germany overwhelmingly the destination.
• What they became. Parish priest and lawyer give way to civil engineer, tracing Norway's move from a pre-industrial to an industrial society.

Built on our own hardware

The whole corpus was processed on SDU's own GPU server — roughly five days of continuous computation across three OCR engines, structured extraction and more than 112,000 adjudication decisions. Total external compute cost: about fifty dollars, for the one phase that used a commercial API.

That number is the argument for on-premises infrastructure in a sentence. Once the hardware exists, work at this scale is effectively free, the source material never leaves the university, and the cost of asking a further question of the corpus is close to zero.

The collaboration

HCNC brings together the Norwegian institutions that hold and study these sources with the transcription expertise at SDU. Hilde Leikny Sommerseth and Bjørn-Richard Pedersen at UiT work on Norway's historical population registers and on the practical problem of reviewing machine transcriptions efficiently; Kristin Ranestad and Paul Sharp have written the source-critical assessment of the Studentene fra… collection itself, establishing what it can and cannot be asked; Nina Boberg-Fazlić brings the economic history of Nordic development.

That combination is deliberate. A digitised source without a source criticism is a trap — this one covers only the university-educated, and its editors handled politically sensitive biographies selectively. Knowing that in advance is what turns 90,028 records into evidence.