The Trails port of WORD-DRIFT — a live knowledge-graph API for lexical semantic change, served from an Oxigraph triple store with SPARQL, provenance, and SHACL validation built in.
WORD-DRIFT is an open knowledge graph modelling how words shift meaning over time — Querdenker from praise to slur, funk from bad smell to musical style. It documents typed drift events (pejoration, amelioration, broadening, narrowing, reappropriation …) and, uniquely, the real-world trigger events that caused each shift. These shifts are not mere lexicographic curiosities: if the words available in a language actively shape what speakers can think and perceive, then each drift event is a datable moment at which the shared cognitive scaffold of a community changes. The Tower of Babel is the mythological limit case — a shared vocabulary fractures and coordination collapses. Drift events are the small, localised Babel-moments that accumulate in living languages.
The original repo pre-generates static JSON files for the browser explorer and validates the RDF offline. word-drift-on-trails replaces that pipeline with a single Trails application that:
- Loads all TTL data into a persistent Oxigraph store on startup
- Exposes the same graph JSON the frontend expects via live SPARQL queries
- Adds a full
/api/sparqlSPARQL endpoint, health check, and provenance tracking - Applies SHACL validation on every write — not just as a one-off CLI script
The static site (site/) is served unchanged. The three JSON endpoints (/graph.json, /graph-core.json, /graph-detail.json) are overridden by live routes registered before the static file mount, so the bundled JSON files act only as an offline fallback.
| Feature | word-drift (original) | word-drift-on-trails |
|---|---|---|
| Storage | rdflib in-memory | Oxigraph persistent |
| Server | python -m http.server |
Trails HTTP adapter (FastAPI) |
| Data format | Pre-generated JSON files | Live SPARQL queries |
| Authentication | None | Bearer token (configurable) |
| Rate limiting | None | Built-in (60 req/min) |
| Health check | None | /api/health endpoint |
| SPARQL endpoint | None | /api/sparql live endpoint |
| Provenance | None | Trails provenance tracking |
| SHACL validation | One-off CLI script | Live on write |
| Versioning | Git-only | KG time-travel |
| Lines of code | ~800 (export.py + serve.sh) | ~400 (Trails handles the rest) |
app.py Trails application — routes + startup
loader.py Loads all TTL files into Oxigraph
graph_builder.py Runs SPARQL → builds graph JSON for the frontend
site/ Static frontend (D3 explorer, unchanged from original)
index.html Landing page / timeline view
explore.html Force-directed graph explorer
about.html Project info
graph.json Bundled fallback (overridden by live /graph.json route)
graph-core.json Bundled fallback (overridden by live /graph-core.json route)
graph-detail.json Bundled fallback (overridden by live /graph-detail.json route)
ontology/ Six Turtle modules (drift: vocabulary)
shapes/ SHACL shape files
examples/ ~200 curated words with drift events and causal hypotheses
data/ ETL output (real, wugs, gfds, alignment, freq, semeval …)
git clone https://github.com/XORwell/word-drift
cd word-drift
# Install Trails + deps
make install
# Run the dev server (auto-reloads, loads data on startup)
make dev
# → http://localhost:8080docker compose up --build
# → http://localhost:8080
# The persistent Oxigraph store lives in the wd_data Docker volume.The first startup loads all TTL files into Oxigraph — expect 10–30 seconds. Subsequent starts reuse the persisted store (fast).
| Method | Path | Description |
|---|---|---|
| GET | / |
Static frontend (index.html) |
| GET | /graph.json |
Full graph for D3 explorer (live SPARQL) |
| GET | /graph-core.json |
Core nodes only — fast first paint |
| GET | /graph-detail.json |
Full detail including causal hypotheses |
| GET | /api/health |
Health check — returns {"status":"ok","triples":N} |
| GET | /api/sparql?query=… |
Live SPARQL 1.1 SELECT/CONSTRUCT endpoint |
| POST | /api/sparql |
SPARQL endpoint (application/sparql-query body) |
| GET | /api/words |
List all words in the KG |
| GET | /api/word/{lemma} |
Full detail for one word |
curl 'http://localhost:8080/api/sparql?query=SELECT+%3Fw+WHERE+%7B+%3Fw+a+drift%3AWord+%7D+LIMIT+10'The drift: vocabulary (in ontology/) covers:
- Lexical — Word / Sense, aligned with OntoLex-Lemon
- Sense over time — attestation intervals, connotation, frequency observations
- Drift event — reified change event + SKOS type taxonomy (valence, scope, mechanism, social pattern)
- Causation —
drift:TriggerEvent(dateable, Wikidata-linkable) - Provenance — PROV-O based, source citation enforced by SHACL
- Causal evidence —
drift:CausalHypothesis: causality as a graded, evidenced hypothesis (not asserted fact)
| Variable | Default | Description |
|---|---|---|
PORT |
8080 |
HTTP listen port |
WORD_DRIFT_STORE |
./wd-store |
Oxigraph persistent store path (:memory: for ephemeral) |
TRAILS_ENV |
development |
development or production |
TRAILS_TOKEN |
(unset) | Bearer token for write endpoints (optional) |
word-drift-on-trails is developed alongside a research paper demonstrating how the Trails framework reduces the infrastructure burden for knowledge-graph applications. The comparison table above is the central empirical claim: the same functionality (persistent store, live SPARQL, SHACL, provenance, HTTP API) requires roughly half the code when built on Trails instead of bespoke glue scripts.
Original WORD-DRIFT paper and data: https://github.com/XORwell/word-drift
Trails framework: https://github.com/XORwell/trails
Code: MIT. Data: see LICENSE-DATA (Creative Commons BY 4.0 for curated examples; corpus-derived data may carry additional restrictions — see data/ subdirectory READMEs).
The core idea for WORD-DRIFT — modelling not just that words shift but why, with causal triggers as first-class entities — was inspired by Lera Boroditsky's TED talk How Language Shapes the Way We Think, which vividly illustrates how deeply language and thought co-evolve over time and across cultures. The talk makes a second point equally central to this project: the words a language makes available actively shape what speakers can think and perceive. Tracking when a word enters or leaves a lexicon, gains or loses a sense, or shifts connotation is therefore not just lexicography — it is a record of changing cognitive possibility.