# Semantic game data attribution

The French semantic rankings are an adaptation by Dictée en ligne of:

- **fastText French Common Crawl + Wikipedia vectors**, Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin and Tomas Mikolov (2018), *Learning Word Vectors for 157 Languages*, LREC 2018. [Paper](https://arxiv.org/abs/1802.06893), [official distribution and licence](https://fasttext.cc/docs/en/crawl-vectors.html). Original vectors: **CC BY-SA 3.0 Unported**, with [legal text](CC-BY-SA-3.0.txt).
- **Lexique 3.83**, Boris New, Christophe Pallier and contributors, **CC BY-SA 4.0**. [Source receipt, adaptation history and attribution](../lexique383/ATTRIBUTION.md), [legal text](../lexique383/LICENSE-CC-BY-SA4.0.txt).

Changes: select eligible lowercase French lemma spellings, apply the documented leisure exclusions and explicit ligature corrections, intersect with the pretrained model's existing vectors, normalize vectors, compute cosine similarity to thirty original targets, quantize display scores, and rank the fixed vocabulary. No model training or runtime subword inference is performed. Full source vectors are not bundled with the application. Source fingerprints and upstream URLs are in [source.json](source.json).

The adapted vocabulary and ranking files are distributed under **CC BY-SA 4.0**. Section 4(b) of the original 3.0 licence permits adaptations under a later licence version with the same elements. Retain attribution, describe further changes, link the licence and share adaptations under its terms. Application logic is separate from these licensed data files. Upstream authors do not endorse this game or its suitability for any individual.

Rankings reflect statistical associations, including ambiguity and source-corpus biases. They are not a dictionary of synonyms, an objective measure of meaning, a cognitive assessment or clinical validation.