Skip to main content

Shared tasks & leaderboard

Created: 07-07-2026 · Last updated: 07-07-2026

Computational lexicography moves fastest where there is a shared task: a fixed public benchmark, a fixed metric, and a leaderboard anyone can join (the model here is the ACL MWSA shared task on monolingual word-sense alignment and its kin). This page hosts the CDSL documentation corpus's own shared tasks: one live, one proposed.

Task 1 — Which-dictionary routing (live)

Problem. Given a one-sentence user scenario ("I met a Sanskritized Pali-looking form in the Mahāvastu, absent from MW"), predict which of the 44 catalogued CDSL dictionaries a working lexicographer would route the user to.

This is the decision every CDSL newcomer faces before any lookup, and the one the which-dictionary quiz teaches. A system that routes well — a rule set, a retrieval system, an LLM prompt — is directly useful as a front-door recommender for the whole corpus.

Benchmark. routing-benchmark.json (v1.0, CC-BY-SA-4.0; card on the data cards page):

  • dev — 18 scenarios (the ones the site quiz also asks; tune freely on these);
  • test — 6 probe scenarios the quiz does not ask, deliberately drawn from the high-novelty tail (SKD/VCP, ACC, GRA, reader vocabularies, compact German);
  • answer space — all 44 catalog codes;
  • metricsstrict accuracy (prediction = gold) and lenient accuracy (prediction in the accepted set: gold + documented defensible alternates). Report both, with the 95% Wilson interval the scorer prints — n is small, the interval is not optional.

Golds are public: this is an open benchmark with a self-reported leaderboard, not a hidden-test competition. The honesty convention is: never train/tune on test, and say in your notes what your system saw.

Run it.

node scripts/score-routing-task.mjs my-predictions.json
# my-predictions.json: { "G-01": "MW", "G-02": "BHS", ..., "P-06": "PW" }

Submit by pull request against csl-guides: add your predictions file under static/shared-task-submissions/ and a leaderboard row below, with the scorer's verbatim output in the PR body.

Leaderboard

Systemdev strictdev lenienttest stricttest lenientDateNotes
Site quiz answer key18/18 = 100% [82.4, 100]100%0/6 = 0%0/6 = 0%2026-07-07The quiz's own key, applied as a router. Perfect on dev by construction (the gold panel re-judged the quiz's scenarios and agreed 18/18 — see the GH-1 caveat); scores 0 on test because it cannot answer scenarios it never asks — exactly the coverage gap GH-1 measured.
Random over 44 codes~2.3% (expected)~3.4% (expected)~2.3% (expected)~4.5% (expected)2026-07-07Analytic expectation, not a run: 1/44 strict; lenient = mean accepted-set size / 44 (dev 27/18 ≈ 1.5 codes, test 12/6 = 2 codes).
your system

Why the baseline rows matter: the gap between the quiz-router's dev and test scores is the benchmark's point. A submission that beats 0% on test while staying strong on dev has learned routing knowledge the site itself does not yet encode — and its test-split error analysis tells us which deep pages or quiz scenarios to write next.

Task 2 — Dictionary-lineage detection (proposed)

Problem. Given a pair of CDSL dictionaries, predict whether a documented derivation relationship exists between them (supplement-of, revision-of, built-on, distillation-of) — e.g. SCH supplements PW; MW built on WIL; PW distills PWG; AP revises AP90.

Why it matters. Lineage is the backbone of the corpus's history (see Origins) and of the headword-overlap cladogram — but the cladogram is similarity, not lineage, and telling descent from mere overlap is a genuine research problem (the same copy-detection question csl-corrections studies via shared-error loci).

Status: proposed — not yet scaffolded. An honest benchmark needs:

  1. a gold edge list of documented derivation relationships, extracted from the 44 deep pages and dictionary front matter (agent-doable here);
  2. a similarity feature layer that is not the gold's own source — the csl-atlas headword-overlap matrix and, ideally, its shared-error loci artifact (csl-atlas-side; consumed when committed, per this site's vendoring rule).

Until (2) exists as a committed atlas artifact, the task stays a proposal rather than a leaderboard with one degenerate submission. Interest / discussion → the csl-guides issue tracker.


Provenance. Tasks designed 07-07-2026 by Fable 5 (claude-fable-5) under handoff H281 (Stream 4, ACL-standards uplift). Task-1 numbers are reproducible from routing-benchmark.json and the committed scorer.

Dr. Mārcis Gasūns