DINOH — a Digital, AI-assisted Infrastructure for Oral History

Within oral history research, recorded life stories are gathered across the Greater Region and its borders. To study them at scale, DINOH provides an AI-assisted pipeline infrastructure in which AI helps read and code interviews — and in which data-protection rules are built into the pipeline and made independently verifiable, rather than left to good intentions. The aim is simple: make oral histories easier to analyse and reuse, while keeping the people in them protected.

Access is conditional

Access is not automatic. Before a researcher may use the pipeline, they must pass a data-protection (DPO) and research-ethics review and receive written approval. Nobody simply runs an interview through it — the review comes first.

Four working parts: Admission, Analysis, Benchmark, Provenance

It works like an assembly line with a guard at the door. Admission is the guard: it classifies every dataset and checks it before it may be used. Analysis is the line itself — it splits interviews into segments and supports AI-assisted coding; in the public demonstration this coding is shown through a transparent, rule-based illustration, while integration of local, open-weight AI tools is gated and in progress. Benchmark is a test bench that measures how good the AI’s suggestions are, using only synthetic data. And Provenance records how each interpretation was made, so findings can always be traced back.

Under the hood — how the four parts work, in plain terms and in code

Admission — the door check, with GDPR built in. Before any interview is analysed, the system checks the dataset the way a data-protection review would: is the file complete and well-formed, are consent and purpose in order, and does processing stay on approved local/EU infrastructure with its privacy status cleared? If anything is missing, the door stays shut — the default answer is “no”, and every refusal is logged with a reason. In engineering terms this is a default-deny gate: the legal rules are checked by code, before data is touched — which is exactly what lets anyone who must respect GDPR and legal status rely on it.
Analysis — reading and coding, with receipts. The interview is split into time-stamped sections, each given thematic labels — and every label must point to the exact words that justify it. In engineering terms this is span-binding: a code with no anchor in the text is rejected. The local AI coding runs primarily on Mistral, an open-weight model held and run on University infrastructure — no interview data leaves the machine.
Benchmark — the practice-run scoreboard. The AI’s suggestions are compared against a human reference and scored — metadata by exact agreement, segmentation by two standard measures, WindowDiff and Pk — always language by language, and only on synthetic data.
Provenance — the tamper-evident logbook. Every run records what went in and what produced it, sealed with a cryptographic fingerprint (SHA-256), so any result can be re-computed and checked rather than taken on trust.

A traffic light for data

Not all data is equally sensitive, so every dataset is given a colour. Green is safely anonymous and can be used right away. Yellow still points to real people — even when names are replaced by codes (pseudonymised) — so it needs a formal data-protection review and a sign-off before any analysis. Red is too sensitive for this route (for example interviews with minors, or about health) and is set aside for separate, case-by-case handling.

Gates that default to “no”

Before any interview reaches the analysis engine, the run must first declare and attest how it will be processed — route, purpose and wording. Each record then passes three automatic checks: the file is well-formed, consent and purpose are in order, and processing stays on approved local/EU infrastructure with its privacy status cleared. If anything required is missing, the gate simply refuses — in engineering terms it is default-deny: the answer is “no” unless everything is in order. Nothing slips through quietly; every refusal is recorded with a reason.

Multilingual by design

The Greater Region is multilingual, and so is DINOH. The benchmark reports on seven languages — German, English, French, Italian, Spanish, Luxembourgish and Portuguese — each judged on its own terms and never merged into one score, so weaker performance in a smaller language cannot hide behind a stronger one.

What the machine does — and what the human in the loop decides

The computer is strict, but it is not in charge. It applies the rules in code and prepares the material — validating records, pseudonymising, keeping everything on approved infrastructure and logging every step. The judgements that matter stay with a human in the loop: whether data is genuinely anonymous, on what legal basis it may be used, how to review the text for anything that could re-identify someone, and how to interpret and code the interviews. And depending on the qualitative-analysis (QDA) method, the researcher can inspect and revise not only the interpretations but the evidence paths that lead to them. Responsibility remains human and is made visible, not hidden behind the technology.

Openness & credit

The evaluation part is openly available and citable: · . The governance engine is described in the concept note. Infrastructure design & engineering: Klaus Behnam Shad (DINOH PI / operator), developed within the LIFE project at the C²DH, ��Vlog. This tab explains methods and safeguards, and is not legal advice.