[DanNet logo]
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.
Table of Contents
Dataset Formats
DanNet is available in multiple formats to maximise compatibility:
Example: Using DanNet with Python
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:
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 representation]
URI Prefixes
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.
[Claude Desktop setup]
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 Theophrastus, our fork of 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):
- Load and clean the previous version's RDF data
- Convert to triples using the current schema
- Import into Apache Jena graphs and apply release changes (only when cutting a release, i.e. when
to differs from from)
- Infer additional triples via OWL/RDFS schemas
- 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.
The rest of ./bootstrap/other/ holds files curated by hand, which aren't downloaded. If the COR files (cor*.tsv, *.cordiff and cor.sem.*.tsv) go missing, you can download them again from ordregister.dk.
If a database build fails partway, delete the half-built db/tdb2/<name> directory before you start again; the build error gives its path. A build only checks that the directory exists, so it would otherwise reuse the incomplete database.
Setup
DanNet requires Java and Clojure's official CLI tools. Dependencies are specified in deps.edn.
Development
- Start an nREPL on port 7888 with
clojure -M:nrepl
- Start the web service using
(restart) in dk.cst.dannet.web.service — available at localhost:3456
- Run the frontend with shadow-cljs:
It serves the app at
localhost:7777, in front of the web service.
Most changes can be tried in the REPL without restarting the web service. Run the tests with clojure -X:validate:test, where the :validate alias adds the validators that the DMLex tests need.
Testing a Release Build
Using Docker (requires Docker daemon running):
Or manually:
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
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:
Deployment
The production server at wordnet.dk runs as a systemd service delegating to Docker.
Service Setup
Updating the Web Service
To update the web service software without changing the database:
Database Release Workflow
When releasing a new version of the database:
-
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.
-
Build the database via REPL in dk.cst.dannet.web.service:
-
Generate the export artifacts. The web service loads only some of the
export namespaces, so require them all first:
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).
-
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.
-
Compact the database, then zip it on the dev machine, ready for transfer:
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.
-
Stop the service on production:
-
Transfer database and export files via SFTP, then:
-
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):
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.
-
Restart:
-
Bump from to the new version and set to to "SNAPSHOT". Clear out the
release-specific block in make-release-changes!: its changes have now
shipped. This readies the next cycle.