babylonhealth/fastText_multilingual

Provides pre-computed SVD-based alignment matrices that project fastText's monolingual word vectors from 78 languages into a shared vector space, enabling cross-lingual similarity and translation prediction via nearest-neighbor lookup. Each language matrix is learned by aligning against English using bilingual dictionaries derived from Google Translate, achieving ~73% precision@1 for translation retrieval while preserving original monolingual relationships. The approach requires only applying a linear transformation to existing fastText vectors—no retraining needed.

1,202 stars. No commits in the last 6 months.