| accretion |
How the record grows. Each stage adds its own block and passes every other block on unchanged (The document contract). |
| actionable workflow |
A workflow on the SSH Open Marketplace that can also be run. In ATRIUM it runs as a Galaxy workflow. |
| AIS CR |
The Archaeological Information System of the Czech Republic, the infrastructure behind AMČR and TEATER, at aiscr.cz (more). |
| ALTO |
An XML format for OCR output. It keeps every word with its position on the page image (more). |
| AMČR |
The Archaeological Map of the Czech Republic (Archeologická mapa České republiky): the national archaeological database of ARÚP and ARÚB, and the partner that runs the ATRIUM pilot. |
| ARÚP, ARÚB |
The Institutes of Archaeology of the Czech Academy of Sciences in Prague and in Brno. They hold the collections the tools were built for (more). |
| block |
One top-level part of the record, written by exactly one tool: page_categories, lines, translations and so on. |
| born-digital |
A document created on a computer, so its text is already stored in the file: a PDF exported from Word, for example. The opposite of a scan. digital-convert reads these. |
| category (page) |
One of the eleven page types page-classification assigns, such as TEXT_T (typed text) or PHOTO (the definitions). |
| CNEC 2.0 |
The entity types of the Czech Named Entity Corpus, used by NameTag's Czech model. nlp-enrich turns its codes into readable names, such as "Geographical name". |
| CoNLL-U |
A plain-text table with one word per line: its lemma, part of speech, grammatical features and syntax (more). |
| CUBBITT |
The Czech–English translation model behind LINDAT's translation service, the translator's default (more). |
DOC_LINE_CATEG/ |
ocr-postprocess's per-line table: every text line with its line category. nlp-enrich reads it. |
| end-to-end test (E2E) |
The test that runs the tools one after another on sample documents and checks what each passes on (W6 on Pipelines). |
| ensemble |
page-classification's five vX.4 models scoring the same page, with their scores averaged. It is the service's default. It has no held-out accuracy figure of its own. |
| FAIR |
Findable, Accessible, Interoperable, Reusable: the principles research-data publication aims at. RO-Crate export is ATRIUM's route to them. |
| fold |
One of five ways the annotated pages were divided into training (80 %), development (10 %) and test (10 %) parts. Each published page-classification model was trained on one fold. |
| GraphQL |
A query language for web APIs, in which the client names exactly the fields it wants. TEATER is queried this way. |
| held-out |
Pages, or other data, a model never saw in training. Accuracy is only honest when it is measured on held-out data. |
| image (container) |
A packaged program, published on GitHub's container registry. Each tool ships at least two: a batch image for the command line and an -api image for the service (Operations). |
| LINDAT/CLARIAH-CZ |
The Czech research infrastructure for language data. It hosts the services nlp-enrich and the translator call: UDPipe, NameTag and translation (more). |
| line category |
The label ocr-postprocess gives each text line: Clear, Noisy, Trash (illegible), Non-text or Empty. |
| NameTag |
LINDAT's named-entity recogniser. nlp-enrich uses NameTag 3 to find names of people, places and organisations. |
| NMT |
Neural machine translation: a model trained to translate whole sentences, such as CUBBITT. |
| OCR / HTR |
Optical character recognition (printed or typed text) and handwritten text recognition. They are different engines, which is why pages are classified first. |
| OpenAPI |
A machine-readable description of a service's endpoints. Each tool commits it as service/openapi.json; it is the API's one contract. |
| originator |
The tool that creates a record's page and line structure: ocr-postprocess for OCR output, digital-convert for born-digital files. |
PAGE_ALTO/ |
ocr-postprocess's output folder: one ALTO file per page, <doc>/<doc>-N.alto.xml. The translator reads it. |
| paradata |
Data about the process: which program, version and settings produced an output, when, and with which components and licences. One JSON file per run (Licences and paradata). |
| pilot (S2) |
AMČR colleagues running the six services themselves, December 2026 – January 2027 (S2 pilot handover). |
| provenance |
The record's account of who contributed what: a stamp on each block, and a list of the runs that wrote to it. |
| quality score |
A number from 0 to 1 that ocr-postprocess gives each line, built from nine measurements of its text. The line category is decided from it and a short list of rules. |
| record |
The ATRIUM document record: one JSON file per document, <doc_id>.document.json, that every stage reads and adds to (The document contract). |
| revision (model) |
A named version of a model on the Hugging Face Hub, such as v4.4. Name the revision you mean: the Hub's default can point to an older one. |
| RO-Crate |
A standard way to package research data with its description, so a repository can publish it (RO-Crate export). |
| routing |
Deciding, from a page's category, which processing it needs next: OCR, handwriting recognition, table extraction, image handling or nothing. A person or a script decides; the tools do not. |
| runbook |
A service's service/README.md: how to start it, what it downloads, what its errors mean and what to do, and its known limits. |
| SKOS |
A W3C standard for publishing vocabularies. ATRIUM's label lists, such as the page categories, are published as SKOS concepts (SKOS). |
| SSH Open Marketplace |
The European catalogue of humanities tools and workflows. Each ATRIUM tool has a record there (more). |
| TEATER |
A thesaurus of archaeological terms with Czech and English names, published by AIS CR. keyword-extract's controlled keywords come from it and from AMČR's keyword lists (more). |
| TEI, TEITOK |
TEI is the standard XML for encoding texts in the humanities. TEITOK is a corpus format and platform built on it, keeping every word linked to its place on the page. nlp-enrich writes TEITOK. |
| top-1, top-3 |
The share of pages whose correct category is the model's first answer (top-1), or among its first three (top-3). |
| UDPipe |
LINDAT's tool for words, lemmas, parts of speech and sentence structure. nlp-enrich uses UDPipe 2 (more). |
| vocabulary (controlled) |
A fixed list of terms that a keyword must come from: here, AMČR's keyword lists and TEATER. A controlled keyword is one of these terms; a statistical keyword is a word taken from the text. |
| workflow narrative |
A tool's workflow told step by step, with the formats in and out of each step. Its SSH Open Marketplace and Galaxy records are built from it (Workflows). |