Sense coarsening

What is sense coarsening?

A coarsening is a map from WordNet 3.0 sense keys to broader concepts. Several WordNet senses of a word can map to one concept, so a prediction is coarse-correct when its WordNet sense lands in the same concept as a gold sense — even if it is not the exact fine sense. Coarsening matters for two reasons:

• Many WordNet distinctions are redundant for applications. Telling a dictionary-construction or language-learning system that volume means "the amount of 3-D space" rather than "a relative amount" rarely matters; telling those apart from "the loudness of a sound" does.
• Fine senses are often not reliably separable. On the hard, human-reviewed lexEN items, expert annotators frequently disagree on the exact fine sense yet agree on the coarse concept — so fine-grained scoring penalises models for distinctions the experts themselves do not make consistently.

How coarsening affects WSD results

• Coarse accuracy is always at least the fine accuracy. Every exact (fine) hit is also a coarse hit, and coarsening can only turn some fine misses into hits, never the reverse. So switching the granularity selector up can only raise a run's score.
• The lift is substantial. On the full lexEN-v1 set, frontier models score around 95% fine-grained and around 98% under Glite coarse. On the hard reviewed subset the gap is much larger — roughly 66% fine versus 78–88% coarse depending on the inventory — which shows that most remaining "errors" are fine sense splits, not genuine disambiguation failures.
• Rankings are largely preserved. Coarsening lifts strong and weak systems together, so the order of models on the leaderboard changes little when you switch granularity.
• It is computed live, from raw predictions. Coarse scores are not baked into a run; SenseBench re-maps each model's predicted sense key through the chosen concept map at build time. Use the granularity selector on the leaderboard to switch between fine, Glite, and CSI; each run page reports all schemes together. See label schemes for how granularity combines with the gold-label selector.

Examples

The four words below each have five or six WordNet senses. In every case Glite groups them exactly the way the consensus of five professional dictionaries (Merriam-Webster, Collins, Cambridge, Oxford Learner's, Longman) does. Within a Glite group, a fine-grained miss still counts as a coarse hit; crossing a group boundary is wrong at both levels. The right-hand note shows where CSI — a domain-based inventory — groups the senses differently.

party (noun) — 5 senses

Glite and the dictionaries merge the social group and the social occasion into one everyday "party" and keep political, search-party, and legal apart. CSI groups them the same way here.

volume (noun) — 6 senses

Glite and the dictionaries merge the two amount senses and the two book senses. CSI instead pairs "3-D space" with "bulk" (both size) and splits the books apart.

care (verb) — 5 senses

Glite and the dictionaries pair "feel concern" with "be concerned with", and "provide care for" with "be in charge of". CSI instead lumps the "wish" sense in with the concern senses.

replacement (noun) — 6 senses

Glite and the dictionaries make one clean cut — the act of replacing vs. the person or thing that does the replacing. CSI scatters the six senses across five domains.

These are typical: across 100 polysemous words, Glite reproduces the professional-dictionary grouping more closely than any other public inventory. Coarse inventories differ mainly in how aggressively and along which lines they merge — which is why SenseBench offers more than one and reports them side by side.

Glite coarse-grained

Glite is a learner-dictionary sense inventory built by Glite. Its concepts are the entries of the Glite Dictionary — a dictionary designed for English learners, where senses are grouped at the granularity a reader actually needs. Because the groupings come from a real, externally published dictionary rather than an automatic rule, Glite's coarsening is principled: across 100 polysemous words it reproduces the sense grouping of major professional dictionaries more closely than any other public inventory.

• How it works. A vendored map sends each WordNet 3.0 sense key to a Glite concept id (for example ct:ct7gsDQ2G22wWJNkuNk4fplan). Senses of a word that the Glite Dictionary treats as one meaning share a concept id; a prediction is coarse-correct when its sense key maps to the same Glite concept as a gold sense.
• Coverage. The public map covers about 95% of the WordNet sense keys lexEN references. A sense key with no Glite concept is scored as its own singleton class (unmapped:<key>) and is never silently merged, so incomplete coverage can only lower coarse accuracy, never inflate it.
• Role. Glite coarse is the default coarse scheme and pairs with the lexEN gold, whose review used the same Glite groupings. The map is vendored from the lexEN release.

Learn more at the Glite Dictionary and glite.ai.

CSI coarse-grained (Lacerra 2020)

CSI — the Coarse Sense Inventory of Lacerra, Bevilacqua, Pasini & Navigli (AAAI 2020) — is an independent, third-party coarsening that groups every WordNet synset into one of 45 broad semantic domains (for example biology, law & crime, art, architecture & archaeology). It is included so that SenseBench's coarse results can be audited against an inventory that the dataset's authors do not control.

• How it works. The released wn_synset2csi mapping assigns each WordNet 3.0 synset to its CSI domain(s). SenseBench vendors a sense-key → concept map derived from it: a synset with several domains is reduced to one composite label (csi:<sorted domains>), a conservative single-class partition, so it plugs into the same scoring contract as Glite with the CSI map swapped in.
• Coverage. CSI maps about 79% of the referenced WordNet keys (uncovered keys become unmapped: singletons). Because its 45 domains are broader than Glite's dictionary concepts, CSI tends to split a word into more classes overall but groups cross-topic senses differently.
• License. CSI is distributed under CC BY-NC-SA 4.0; the vendored subset keeps the same terms with attribution to the original authors.