README.md

floret: fastText + Bloom embeddings for compact, full-coverage vectors with spaCy

floret is an extended version of fastText that can produce word representations for any word from a compact vector table. It combines:

• fastText's subwords to provide embeddings for any word
• Bloom embeddings ("hashing trick") for a compact vector table

Installation

pip install floret

Usage

Train floret vectors using the options:

• mode: "floret", storing both words and subwords in the same compact hash table
• hashCount: store each entry in 1-4 rows in the hash table (recommended: 2)
• bucket: in combination with hashCount>1, the size of the hash table can be greatly reduced (recommended: 25000--100000, reduced from the fastText default of 2000000)
• minn: min length of char ngram (default: 3)
• maxn: max length of char ngram (default: 6)

import floret

# train vectors
model = floret.train_unsupervised(
    "data.txt",
    model="cbow",
    mode="floret",
    hashCount=2,
    bucket=50000,
    minn=3,
    maxn=6,
)

# query vector
model.get_word_vector("broccoli")

# save full model
model.save_model("vectors.bin")

# export standard word-only vector table
model.save_vectors("vectors.vec")

# export floret vector table
model.save_floret_vectors("vectors.floret")

Note: with the default setting mode="fasttext", floret trains original fastText vectors.

Use floret vectors in spaCy

Import floret vectors into spaCy v3.2+:

spacy init vectors LANG vectors.floret spacy_vectors_model --mode floret

Notes

floret contains all features of the original fasttext module. See the fasttext docs for more information.

The fasttext and floret binary formats saved with model.save_model("model.bin") are not compatible.