WISSE — Sentence embeddings

Sentence embeddings via entropy-weighted series (TF-IDF–weighted word embeddings). No language or knowledge resources required. Python 3.8+.

SBERT-like API: encode() and similarity(). Default model keys (wisse-fasttext-300, wisse-idf-en) point to the Hugging Face repo; once the paper’s Wikipedia FastText and TF-IDF assets are uploaded there (one-time, see Uploading assets), they auto-download on first use to ~/.wisse. Until then, use local paths (e.g. after downloading from MEGA).

---

Quick start

With assets on Hugging Face (after one-time upload of the paper assets):

from wisse import SentenceEmbedding

model = SentenceEmbedding()  # downloads to ~/.wisse on first use

sentences = [
    "The weather is lovely today.",
    "It's so sunny outside!",
    "He drove to the stadium.",
]
embeddings = model.encode(sentences)  # shape (3, 300)
sim = model.similarity(embeddings, embeddings)

With local paths (e.g. after downloading from MEGA):

model = SentenceEmbedding(
    model_name_or_path="/path/to/indexed_fasttext/",
    idf_name_or_path="/path/to/idf-en.pkl",
)
embeddings = model.encode(["First sentence.", "Second sentence."])

Similarity: "cosine", "dot", "euclidean", "manhattan".

from wisse import similarity
import numpy as np
s = similarity(embeddings, embeddings, similarity_fn="cosine")

---

Installation

pip install -e .

Requirements: Python ≥3.8, numpy, scikit-learn, gensim, joblib, requests.

---

Default models (Hugging Face)

The package expects Wikipedia-trained FastText (300d) and TF-IDF weights at the Hugging Face repo. Once those files are uploaded (see Uploading assets), they download on first use to ~/.wisse (or $WISSE_HOME).

• Registry keys: wisse-fasttext-300, wisse-idf-en (and optionally wisse-glove-300).
• Override: WISSE_HF_REPO, WISSE_HF_REPO_TYPE, or WISSE_FASTTEXT_URL, WISSE_IDF_URL, WISSE_EMBEDDING_URL.

---

Repository layout

Clean Python package layout:

sentence_embedding/
├── wisse/                 # Package
│   ├── __init__.py
│   ├── wisse.py           # Core: TF-IDF weighting, vector_space, keyed2indexed
│   ├── model.py           # SentenceEmbedding (SBERT-like API)
│   ├── similarity.py      # Pairwise similarity helpers
│   ├── download.py        # HF registry and autodownload
│   └── cli.py             # wisse-encode, keyed2indexed entry points
├── tests/                 # Pytest suite
├── hf_model/              # Hugging Face model card (README.md) and push script
├── setup.py
├── pyproject.toml
├── requirements.txt
├── run_tests.py
├── keyed2indexed.py      # Standalone script (or use CLI after install)
├── LICENSE
├── .gitignore
└── README.md

---

CLI

After pip install:

wisse-encode --input sentences.txt --output vectors.npy
wisse-encode --input sentences.txt --output out.npy --model wisse-fasttext-300 --idf wisse-idf-en

keyed2indexed --input model.bin --output output_indexed
keyed2indexed --input model.vec --txt --output output_indexed

wisse-train --wikipedia en --idf-out idf-en.pkl --embeddings-out fasttext-300-indexed
wisse-train --corpus-dir ./my_texts --document-unit paragraph --idf-out idf.pkl

Train IDF + FastText (new operating mode): from a directory of plain text files or from Wikipedia (Hugging Face). Produces WISSE-ready IDF pickle and indexed FastText embeddings. For Wikipedia you need pip install ".[train]" (adds datasets).

• --corpus-dir PATH — directory of plain text files, or
• --wikipedia LANG — e.g. en, es (downloads from HF wikimedia/wikipedia).
• --document-unit article|paragraph — one doc per file/article vs per paragraph.
• --idf-out, --embeddings-out — explicit output paths (defaults: idf-<lang>.pkl, fasttext-300-indexed).
• --binary-out PATH — optionally save FastText in Word2Vec binary format.
• --dim, --window, --min-count, --epochs — paper defaults (300, 5, 5, 5), all configurable.
• --cap-articles, --cap-tokens — optional cap with efficient random sampling; default for Wikipedia: 500k articles / 100M tokens.

From repo without installing:

python keyed2indexed.py --input model.bin --output output_indexed

---

Low-level usage

Convert word2vec to indexed format and use the WISSE combiner:

import wisse
from gensim.models.keyedvectors import KeyedVectors

kv = KeyedVectors.load_word2vec_format("/path/to/embeddings.bin", binary=True)
wisse.keyed2indexed(kv, "/path/to/output_dir/")

embedding = wisse.vector_space("/path/to/output_dir/")
# embedding["word"] → array

import pickle
with open("/path/to/tfidf.pkl", "rb") as f:
    vectorizer = pickle.load(f)
w = wisse.wisse(embedding, vectorizer=vectorizer, tf_tfidf=True, combiner="sum", generate=True)
vec = w.infer_sentence("this is a sentence")

---

Testing

pip install -e ".[dev]"
pytest tests/ -v
# or
python run_tests.py

Test	Description	Needs real pretrained assets?
test_00_install	Package and public API import	No
test_01_toy_tfidf_fasttext_and_helpers	Toy TF-IDF FastText + all helpers (synthetic data)	No
test_02_paper_wikipedia_assets	Paper’s Wikipedia FastText + TF-IDF	Yes (see below)
test_new_user_full_workflow	New user: mocked download, toy sentences	No (uses mocks)
test_encode_similarity	encode/similarity and low-level API (synthetic)	No

Tests that use the real pretrained models (optional): test_02_paper_wikipedia_assets runs only when assets are available:

• Set WISSE_PAPER_FASTTEXT_DIR and WISSE_PAPER_IDF_PATH to local paths (e.g. after downloading from the MEGA links below), or
• Set WISSE_TEST_HF_REGISTRY=1 to use the Hugging Face registry (only works after the paper assets have been uploaded to the HF repo).

---

Uploading assets to Hugging Face

For SentenceEmbedding() to work with defaults (no local paths), the paper’s FastText and TF-IDF files must be on the Hub. There are no synthetic “minimal” defaults — only the real pretrained assets are useful.

1. Get the assets: Download from MEGA (links in Pretrained assets): indexed FastText and the TF-IDF pickle. Optionally pack the FastText directory as fasttext-300-indexed.tar.gz.
2. Upload: Use hf_model/upload_assets_to_hf.py with local paths, or the Hub UI/CLI. Full steps: hf_model/PUSH_TO_HF.md.
3. Repo: huggingface.co/datasets/iarroyof/wisse-models

---

Pretrained assets (manual download)

If you prefer not to use the Hub, download and extract manually, then pass paths to SentenceEmbedding(...) or wisse.vector_space(path):

• FastText (Wikipedia 300d): idx_FastText
• Word2Vec (Wikipedia 300d): idx_Word2Vec
• GloVe (840B 300d): idx_Glove
• Dep2Vec (300d): idx_Dep2Vec
• TF-IDF (Wikipedia, stop words ignored): pretrained_idf

---

Citation

@article{arroyo2017unsupervised,
  title={Unsupervised Sentence Representations as Word Information Series: Revisiting TF--IDF},
  author={Arroyo-Fern{\'a}ndez, Ignacio and M{\'e}ndez-Cruz, Carlos-Francisco and Sierra, Gerardo and Torres-Moreno, Juan-Manuel and Sidorov, Grigori},
  journal={arXiv preprint arXiv:1710.06524},
  year={2017}
}