babelvec 0.1.0

Position-aware, cross-lingually aligned word embeddings built on FastText

BabelVec

Position-aware, cross-lingually aligned word embeddings built on FastText.

Features

• Position-Aware Embeddings: Word order matters! Uses RoPE, sinusoidal, or decay positional encoding
• Cross-Lingual Alignment: Ensemble alignment (Procrustes + InfoNCE) for multilingual compatibility
• FastText Foundation: Handles OOV words through subword information
• Multiple Training Modes: Monolingual, multilingual, or post-hoc alignment

Installation

pip install babelvec

For visualization support:

pip install babelvec[viz]

Quick Start

from babelvec import BabelVec

# Load a model
model = BabelVec.load('path/to/model.bin')

# Get word vector
vec = model.get_word_vector("hello")

# Position-aware sentence embedding (order matters)
vec1 = model.get_sentence_vector("The dog bites the man", method='rope')
vec2 = model.get_sentence_vector("The man bites the dog", method='rope')
# vec1 != vec2 because word order is different!

# Standard averaging (order-agnostic)
vec_avg = model.get_sentence_vector("The dog bites the man", method='average')

Training

Monolingual Training

from babelvec.training import train_monolingual

model = train_monolingual(
    lang='en',
    corpus_path='corpus.txt',
    dim=300,
    epochs=5
)
model.save('en_300d.bin')

Multilingual Training with Alignment

from babelvec.training import train_multilingual

model = train_multilingual(
    languages=['en', 'fr', 'de'],
    corpus_paths={'en': 'en.txt', 'fr': 'fr.txt', 'de': 'de.txt'},
    dim=300,
    alignment='ensemble'
)

Post-hoc Alignment

from babelvec.training import align_models

aligned = align_models(
    models={'en': model_en, 'fr': model_fr},
    method='ensemble',
    parallel_data=parallel_sentences
)

Positional Encoding Methods

Method	Description	Use Case
average	Simple averaging (no position)	Bag-of-words tasks
rope	Rotary Position Embedding	Best for semantic similarity
sinusoidal	Transformer-style positional	General purpose
decay	Exponential position decay	Emphasis on early words

Citation

@software{babelvec2025,
  title = {BabelVec: Position-Aware Cross-Lingual Word Embeddings},
  author = {Kamali, Omar},
  year = {2025},
  url = {https://github.com/omarkamali/babelvec}
}

License

MIT License - see LICENSE for details.

Copyright © 2025 Omar Kamali