Superb language data

Meaning you can inspect.

Follow a word to its specific meaning. Compare a claim with the passage behind it. See where a revision changes what someone could understand.

Try an argument OpenAPI specification

Use the public API

Browse the dataset catalogue for coverage, versions and review status.

GET /api/lexicon/words/bank
GET /api/usage/contrasts?channel=claim
GET /api/datasets/literary-context/records
POST /api/argument/review

Word meanings & connections

The core inventory has 127,311 words, 185,129 meanings and 1,227,481 directed connections from Open English WordNet 2025. Eight sourced editorial slang senses extend that inventory. Every connection keeps its target meaning ID and relation type. A river bank and a financial bank remain distinct.

GET /api/lexicon
GET /api/lexicon/words?prefix=ban
GET /api/lexicon/words/bank?sense=oewn-bank__1.17.01..

Explore “significant” for an example of the separate contextual notes. Sixteen selected terms have original illustrative readings and clarification questions. Empty usageNotes.notes means no editorial note is available.

Definitions can be technical or historical. Meaning order does not establish a speaker's intent. Notes are an AI-assisted editorial preview, not independently reviewed usage research.

Contextual contrasts & literary passages

The usage preview contains 12 original situations and 36 contrasts: context, before and after wording, intended effect, explanation, limits and provenance. These AI-assisted teaching examples have not received independent review and are not a validated benchmark.

Compare a reply · Download contrasts · Manifest and hashes

The literary release contains 5,096 attributed historical excerpts. Association practice distinguishes definition crossings, book co-occurrence and unresolved editorial evidence; it is not a validated semantic graph.

GET /api/datasets/literary-context/download
GET /api/datasets/association-practice/records/weak

Argument review & transcripts

Send an argument and up to four source excerpts. Local-mode API checks return related passages with exact text offsets, possible changes in scope or certainty, and contextual notes. Overlapping wording is not a verdict.

{
  "text": "Reading ability improved.",
  "sources": [{
    "id": "report", "label": "Study excerpt",
    "text": "Reading ability was not measured."
  }],
  "mode": "local"
}

Optional cloud review assesses support within the supplied text, competing interpretations and a possible revision. It needs a configured provider, a scoped access key and explicit consent. Check current capabilities. Video transcription returns editable speech text; it does not analyze visual evidence or fetch TikTok links.

Neither mode authenticates sources, independently fact-checks the world, or awards a truth score. API requests are processed without server storage by Superb; cloud provider retention terms still apply. Text checks inside the app run on the device.

Versions, limits & response handling

Dataset pages contain at most 100 records; follow nextCursor. Pin dataset and usage releases with ?version=…. Stale versions or cursors return 409. JSONL downloads include source terms, byte lengths and SHA-256 hashes in their manifests.

Public data uses GET or HEAD and can be cached for an hour. Argument review uses POST and no-store. All API routes support browser preflight. Missing data returns 404, malformed input 400, oversize uploads 413, and unavailable services 503. See OpenAPI for the full contract.

Useful for a specific failure.

For summarization, inspect unsupported additions and lost qualifications. For conversation tools, separate a disagreement about facts from one about definitions or values. For learning products, retain the source and meaning that make an explanation teachable.

A custom evaluation pilot should use newly authored cases, independent review and a private holdout. The public previews demonstrate the format; they do not establish evaluation accuracy or representative coverage. Discuss an evaluation use case.

Sources & rights

Open English WordNet 2025, derived from Princeton WordNet, supplies the dictionary under CC BY 4.0. Keep its attribution when redistributing entries. Superb's original contrasts and contextual notes also use CC BY 4.0. The slang supplement uses original editorial paraphrases with brief third-party usage excerpts. Those quotations retain their source rights and are not part of the CC BY licence or approved training data. Follow the attribution on each entry. Other datasets retain their own terms.

Each literary excerpt retains its source and rights information. Observed-language releases require reviewed records, documented redistribution permission and private access. The book catalogue remains independently accessible; public availability does not erase source rights.