Data Sources & Licenses

1. ConceptNet Numberbatch

• Version: Numberbatch English 19.08.
• Authors and maintainers: Robyn Speer, Joshua Chin, Luminoso Technologies, and the ConceptNet team.
• License: Creative Commons Attribution-ShareAlike 4.0 International.
• License link: CC BY-SA 4.0.
• Source: ConceptNet Numberbatch 19.08 English data.
• How it is used: The vectors provide the semantic similarity scores between clue words and candidate words.
• Modifications: Yes.

The current build selects English entries that intersect with the cleaned ENABLE vocabulary. Word labels are normalized to lowercase. The original 300 dimensions are retained.

Each vector is converted from floating-point values to signed 8-bit values with a per-vector scale and norm. The data is then split into static binary shards.

These changes involve filtering, format conversion, quantization, and sharding. No dimensionality reduction is performed.

2. WordNet

• Version: WordNet 3.0, obtained through the NLTK corpus.
• Owner: Princeton University.
• License: WordNet 3.0 License.
• License link: Princeton University's WordNet license.
• Source: WordNet at Princeton University.
• How it is used: WordNet supplies synonyms, broader and narrower terms, antonyms, lemmas, word forms, parts of speech, and short definitions.

It also supports filtering for plurals, inflected forms, proper names, and words that are too closely related to a clue.

Modifications: Yes.

Selected lexical fields are extracted and exported as JSON. Relation endpoints are normalized to lemmas. References outside the final vocabulary are removed.

Reciprocal relation edges are added where needed. The original WordNet database is not altered.

Required notice: WordNet 3.0 Copyright 2006 by Princeton University. All rights reserved.

3. wordfreq

• Version: wordfreq 3.1.1.
• Author and maintainer: Robyn Speer.
• Code license: Apache License 2.0.
• Data license: Creative Commons Attribution-ShareAlike 4.0 International.
• License links: Apache License 2.0 and CC BY-SA 4.0.
• Source: wordfreq on PyPI.
• How it is used: Zipf frequency scores help filter extremely rare candidates and provide deterministic tie-breaking.

These scores do not replace the semantic similarity score or determine the primary ranking formula.

Modifications: Yes.

Frequency scores are exported for the cleaned ENABLE vocabulary at build time. The exported values are rounded and merged into the website's lexicon.

Candidates below the build's frequency threshold are excluded. The original wordfreq package is not modified.

wordfreq also contains data from upstream sources identified in its attribution notice. The applicable notice is retained with the website's static license files.

4. ENABLE word list

• Version: enable1.txt at commit 233d25a56b9ac8bef906916ccf091a976995cce0.
• Authors and maintainers: M. Cooper and the ENABLE contributors. The pinned copy is hosted in the dolph/dictionary repository.
• Status: Public domain.
• Declaration: ENABLE public-domain declaration.
• Pinned source: ENABLE enable1.txt.
• How it is used: ENABLE provides the base English vocabulary used to validate candidate words.

It does not guarantee that a particular game or game version will accept a word.

Modifications: Yes.

Entries are converted to lowercase. Only entries containing 2–15 English letters are retained.

The resulting vocabulary is intersected with the available Numberbatch vectors and enriched with WordNet and wordfreq data. Additional eligibility filters are applied during candidate selection.