Klaus Behnam Shad

Computational assistance in oral history needs clearly defined tasks, inspectable procedures and human judgement. Software output alone does not establish scholarly validity.

DINOH Evaluation makes metadata and segmentation scoring procedures inspectable using 28 synthetic records in seven languages. The transcript texts were drafted with AI assistance and edited by the author, the reference annotations were written by a researcher. The current model backend is a placeholder, so its task scores say nothing about model performance. OHPIPE is available separately as an experimental preview of the transcript workflow. Required project reviews precede processing, the software grants no authorisation.

MIT License · DINOH Evaluation v1.0.0 archived on Zenodo, later development on GitHub · Archived, all versions: 10.5281/zenodo.21273366 · Governance

Pipeline component of DINOH: a research workflow for oral-history transcripts, descriptive proposals and human review, with a manual pseudonymisation workflow and interfaces for local model adapters. The public preview runs a reproducible demonstration on invented transcripts and recorded model responses; known test failures are documented in the repository.

MIT License · experimental public preview (0.1.0a1) · Release v0.1.0a1

Automatic transcripts sometimes assign the wrong speaker, and in a long interview it is unclear where listening again is worthwhile.

diar2 combines Whisper transcription with Nemotron diarization and assigns a speaker to every word. A review list names the passages where the assignment is uncertain: speaker changes within a sentence, very short turns and overlaps. After installation it runs offline; the self-test repeats the whole pipeline with the network blocked. Accuracy has so far been measured only on a synthetic dialogue; a bundled procedure measures it on hand-annotated excerpts of real interviews.

Apache-2.0 · Public repository · German by default · Apple Silicon only · No DOI yet

Interview transcripts are full of names, places, employers and dates that make people indirectly identifiable. Under the GDPR that material cannot simply be handed to a cloud service for anonymisation, and automatic anonymisation on its own is not a sufficient safeguard either.

AegisQDA runs entirely on local infrastructure. Presidio-backed recognition of personal data is combined with mandatory human review; identifiers are replaced by typed, document-local surrogates and checked again in a second verification pass. Data is released to a downstream analysis tool only after explicit sign-off.

MIT License · v0.1.0 (MVP) · Archived on Zenodo, all versions: 10.5281/zenodo.21854346

Language models produce coding that sounds convincing and cannot be checked. For qualitative research that is worse than no automation at all: an unverifiable claim about a source is not a finding.

DigQDA binds every output to its source. Versioned method contracts fix what is coded and how; prompts stay close to the source unit and return machine-readable JSON; segmentation is deterministic; quotes and locators are validated. A fail-closed conformance suite rejects results that do not resolve to the text.

MIT License · pre-release (v0.4)

Qualitative coding is usually not reproducible. Categories emerge during the work and are rarely recorded in a form a second reader could apply to the same text and arrive at the same result.

Relational Justice Analysis makes the rules explicit up front: a rule-based, fully traceable annotation framework with indicators for German, English and French. Every annotation can be traced back to the rule that produced it.

MIT License · v2.0.0 · Archived on Zenodo, all versions: 10.5281/zenodo.18640259

Interview data sits in incompatible silos. Every project invents its own fields, so material cannot be found, cited or reused across disciplines and repositories.

A deliberately minimal core of thirteen fields, seven of them mandatory, plus extensible implementation profiles for individual disciplines. Specified in DCTAP, with crosswalks to Dublin Core, schema.org/Dataset, REFI-QDA/QDPX and CMDI. Built for interoperability instead of lock-in.

CC BY 4.0 · v1.0 · Archived on Zenodo, all versions: 10.5281/zenodo.20507328

TEI CRM Bridge makes the information in a letter edition usable as linked data; every name mention stays bound to its place in the text. The tool takes over the persons, places and organisations encoded by the editors, together with sending and receiving details, from TEI P5 letters. An NER model for historical German (Impresso) can suggest further names. In the output graph, these remain explicitly candidates. In a local review view, a person can accept, reject or leave open each occurrence; acceptances and rejections are recorded with the reviewer and the time in a separate graph.

The evaluation on 40 Schnitzler letters measures how closely v0.2 agrees with the encoding guidelines of this edition. In paragraph text, F1 was 0.668 strict and 0.747 overlapping. In v0.1, new names were placed at an inadmissible position in 307 of 336 cases; in v0.2, in 0 of 55. The existing encoding was recoverable in all 40 letters. The review function of v0.3 adds a traceable human decision step; it has not been measured separately.

Planned from 0.4.0: CMIF export for exchange with correspSearch and authority matching via PMB. For this release, the measurement environment and the data state will be fixed, queries documented and the figures checked again. Until then, the F1 values given here remain explicitly results of v0.2.

MIT License · v0.3.0 (prototype) · Archived on Zenodo: 10.5281/zenodo.22979658 · Project page with evaluation