dannet

DanNet is a WordNet for the Danish language. DanNet uses RDF as its native representation at both the database level, in the application space, and as its primary serialisation format.

• Browse the data at wordnet.dk
• Query the data by integrating with AI
• Download datasets from the releases page

Table of Contents

• Dataset Formats
• Companion Datasets
• Standards
• LLM Integration (AI)
• Implementation
• Setup
• Deployment
• Database Release Workflow

Dataset Formats

DanNet is available in multiple formats to maximise compatibility:

Format	Description
RDF (Turtle)	Native representation. Load into any RDF graph database (such as Apache Jena) and query with SPARQL.
CSV	Published with column metadata as CSVW.
WN-LMF	XML format compatible with Python libraries like wn.
DMLex	OASIS DMLex 1.0 representation as both XML and JSON, in a Danish and an English variant. The DMLex browser shows it as a dictionary.

Example: Using DanNet with Python

import wn

wn.add("dannet-wn-lmf.xml.gz")

for synset in wn.synsets('kage'):
    print((synset.lexfile() or "?") + ": " + (synset.definition() or "?"))

Differences Between Formats

While every format includes all synsets/senses/words, the CSV, WN-LMF and DMLex variants do not include every data point:

• CSV: Some data is lost when converting from an open graph to fixed tables.
• WN-LMF: Only official GWA relations are included per the standard (proprietary DanNet relations from the DanNet schema are excluded).
• DMLex: Combines DanNet with COR, DDS and COR.SEM in one file. Relations to other datasets are not included. See doc/dmlex/plan.md for the conversion rules.

For the complete dataset, use the RDF format or browse at wordnet.dk.

Companion Datasets

Several companion datasets expand the RDF graph with additional data:

Dataset	Description
COR	Links DanNet resources to IDs from the COR project.
DDS	Adds sentiment data to DanNet resources.
OEWN extension	Provides DanNet-style labels for the Open English WordNet to facilitate browsing connections between the two datasets.

Inferred Data

Additional data is implicitly inferred from the base dataset, companion datasets, and ontological metadata. These inferences can be browsed at wordnet.dk. Releases containing fully inferred graphs are specifically marked as such.

Standards

DanNet is based on the Ontolex-lemon standard combined with relations defined by the Global Wordnet Association as used in the official GWA RDF standard.

Ontolex-lemon class	Represents
ontolex:LexicalConcept	Synsets
ontolex:LexicalSense	Word senses
ontolex:LexicalEntry	Words
ontolex:Form	Forms

URI Prefixes

Prefix	URI	Purpose
dn	https://wordnet.dk/dannet/data/	Dataset instances
dnc	https://wordnet.dk/dannet/concepts/	Ontological type members
dns	https://wordnet.dk/dannet/schema/	Schema definitions
dnf	https://wordnet.dk/dannet/function/	Custom SPARQL functions (dnf:path, dnf:lch, dnf:wup synset similarity)

All DanNet URIs resolve to HTTP resources. Accessing one of these URIs via a GET request returns the data for that resource.

Schemas

DanNet has proprietary relations defined in the DanNet schema in an Ontolex-compatible way. There is also a schema for EuroWordNet concepts. Both schemas follow the RDF conventions listed by Philippe Martin.

LLM Integration

DanNet can be connected to AI tools like Claude via MCP (Model Context Protocol).

• MCP server URL: https://wordnet.dk/mcp
• Registry ID: io.github.kuhumcst/dannet

To connect in e.g. Claude Desktop: go to Settings > Connectors > Browse Connectors, click "add a custom one", enter a name (e.g., "DanNet") and the MCP server URL.

Once connected, you can query DanNet's semantic relations directly through Claude.

Implementation

The database backend is Apache Jena, a mature RDF triplestore with OWL inference support. When represented in Jena, DanNet's relations form a queryable knowledge graph. DanNet is developed in Clojure, using libraries like Aristotle to interact with Jena.

See rationale.md for more on the design decisions.

Full Production Setup

The production deployment at wordnet.dk consists of three services managed via Docker Compose:

• DanNet — the Clojure/ClojureScript web application
• MCP server — a Python-based MCP server providing LLM access to DanNet
• Caddy — reverse proxy handling HTTPS and routing

Clojure Support

DanNet can be queried in various ways from Clojure (see queries.md). Apache Jena transactions are built-in and enable persistence via the TDB 2 layer.

Web Application

The frontend is written in ClojureScript using Rum, served by Pedestal. The app works both as a single-page application (with JavaScript) and as a regular HTML website (without). Content negotiation serves different representations (HTML, RDF, Transit+JSON) based on the request.

See doc/web.md for details.

Bootstrap Process

New releases are bootstrapped from the preceding release. The process (in dk.cst.dannet.db.bootstrap):

1. Load and clean the previous version's RDF data
2. Convert to triples using the current schema
3. Import into Apache Jena graphs and apply release changes (only when cutting a release, i.e. when to differs from from)
4. Infer additional triples via OWL/RDFS schemas
5. Export the final RDF dataset (see Database Release Workflow)

Bootstrap data lives under ./bootstrap relative to the execution directory: the DanNet release assets in ./bootstrap/from/<version>/ (named after the release being bootstrapped from, so several can coexist) and the shared English datasets in ./bootstrap/other/english/. Missing files are downloaded automatically, so manual placement is only needed when working offline.

Setup

DanNet requires Java and Clojure's official CLI tools. Dependencies are specified in deps.edn.

Development

1. Start the web service using (restart) in dk.cst.dannet.web.service — available at localhost:3456
2. Run the frontend with shadow-cljs:

   npx shadow-cljs watch app
   

Testing a Release Build

Using Docker (requires Docker daemon running):

# From the docker/ directory
docker compose up --build

Or manually:

shadow-cljs --aliases :frontend release app
clojure -T:build org.corfield.build/uber :lib dk.cst/dannet :main dk.cst.dannet.web.service :uber-file "\"dannet.jar\""
java -jar -Xmx4g dannet.jar

Memory Requirements

The system uses ~1.5 GB when idle and ~3 GB when rebuilding the database. A server should have at least 4 GB of available RAM.

Validating RDF (SHACL)

The dn: dataset is validated against SHACL shapes located in resources/schemas/internal/shapes/ (see dk.cst.dannet.db.shapes). This happens in several ways:

• a non-fatal check of the asserted graph runs asynchronously at every boot, logging violations and comparing counts to a known baseline,
• RDF exports of the dn: dataset are gated: a baseline regression aborts the export, and
• fixture-based tests run via clojure -X:validate:test, which is also executed by the GitHub Actions workflow in .github/workflows/test.yml.

Validating WN-LMF

python3 -m venv examples/venv
source examples/venv/bin/activate
python3 -m pip install wn
python -m wn validate --output-file examples/wn-lmf-validation.json export/wn-lmf/dannet-wn-lmf.xml

Validating DMLex

The validator in dk.cst.dannet.db.export.dmlex-validate checks both serializations of a variant against the official DMLex schemas. It needs the :validate alias:

clojure -M:validate -e "((requiring-resolve 'dk.cst.dannet.db.export.dmlex-validate/validate-dmlex!) \"export/dmlex/\" \"da\")"

Deployment

The production server at wordnet.dk runs as a systemd service delegating to Docker.

Service Setup

cp system/dannet.service /etc/systemd/system/dannet.service
systemctl enable dannet
systemctl start dannet

Updating the Web Service

To update the web service software without changing the database:

# From the docker/ directory
docker compose up -d dannet --build

Database Release Workflow

When releasing a new version of the database:

1. Set to in dk.cst.dannet.release to the new version, leaving from on the release being bootstrapped from. The release-specific changes in make-release-changes! only run once the two differ.

2. Build the database via REPL in dk.cst.dannet.web.service:

   (restart)
   

3. Generate the export artifacts, each in its own namespace:

   (dk.cst.dannet.db.export.rdf/export-rdf! @dk.cst.dannet.web.resources/db)
   (dk.cst.dannet.db.export.csv/export-csv! @dk.cst.dannet.web.resources/db)
   (dk.cst.dannet.db.export.wn-lmf/export-wn-lmf! "export/wn-lmf/")
   (dk.cst.dannet.db.export.dmlex/export-dmlex-variants! "export/dmlex/" @dk.cst.dannet.web.resources/db)
   ;; ~6 minutes
   (dk.cst.dannet.db.query/save-synset-indegrees!
     (:graph @dk.cst.dannet.web.resources/db))
   

   This writes export/rdf/ (dannet.zip, cor.zip, cor-sem.zip, framenet.zip, dds.zip, oewn-extension.zip), export/csv/dannet-csv.zip, export/wn-lmf/dannet-wn-lmf.xml.gz, export/dmlex/ (dannet-dmlex-da.zip, dannet-dmlex-en.zip) and export/synset-indegree.edn. These ship to production (step 7) and become the GitHub release assets that the next cycle bootstraps from (step 4).

4. Publish a GitHub release tagged v<version> and attach the bootstrap assets listed by bootstrap-files in dk.cst.dannet.db.bootstrap.downloads: dannet.zip, cor.zip, cor-sem.zip, dds.zip, oewn-extension.zip and synset-indegree.edn. The next cycle fetches these from GitHub.

5. Compact the database, then zip it on the dev machine, ready for transfer:

   (dk.cst.dannet.db/compact! (:dataset @dk.cst.dannet.web.instance/db))
   

   TDB2 only reclaims the space left by in-place updates when compacted, and writes a new Data-000N generation, so restart the service afterwards. Before transferring, check that the database size divided by the triple count is in the hundreds of bytes, not the thousands.

6. Stop the service on production:

   docker compose stop dannet
   

7. Transfer database and export files via SFTP, then:

   unzip -o tdb2.zip -d /dannet/db/
   mv cor.zip cor-sem.zip framenet.zip dannet.zip dds.zip oewn-extension.zip /dannet/export/rdf/
   mv dannet-csv.zip /dannet/export/csv/
   mv dannet-wn-lmf.xml.gz /dannet/export/wn-lmf/
   mv dannet-dmlex-da.zip dannet-dmlex-en.zip /dannet/export/dmlex/
   

8. Ship the export/synset-indegree.edn generated in step 3. Production runs with --no-bootstrap and so never downloads it, but it is read at query time to rank search results and entity relations, and it should describe the database actually being shipped. Either location works, the first taking precedence (see indegrees-files in dk.cst.dannet.db.query):

   mv synset-indegree.edn /dannet/db/                      # legacy location
   mv synset-indegree.edn /dannet/bootstrap/from/2026-08-03/   # alongside the bootstrap inputs
   

   If neither exists the service still starts and search still works, but results come back unranked and a :dannet.query/indegrees-unavailable error is logged.

9. Restart:

   docker compose up -d dannet --build
   

10. Bump from to the new version and delete to, which then defaults to from again. Clear out the release-specific block in make-release-changes!: its changes have now shipped. This readies the next cycle.