Hailay/fasttext-tigrinya

FastText Word Embeddings for Tigrinya

Subword-aware word embeddings for Tigrinya (Ge'ez script), trained with FastText. The model provides vector representations for Tigrinya word forms, including out-of-vocabulary words via character n-gram composition.

Tigrinya is a morphologically rich, low-resource language for which few pretrained embedding resources exist. Subword-aware embeddings are particularly suited to such languages, since inflected forms unseen during training can still be represented from their character n-grams.

Technical details

Metadata extracted from the released FastText binary.

Property	Value
Architecture	FastText, skip-gram
Embedding dimension	300
Vocabulary size	156,687
Character n-gram range	2–5
Context window	5
Training epochs	10
Minimum word count	5
Negative samples	5
Loss	negative sampling
Bucket size	2,000,000
Word n-grams	1

Training is not reproducible from this repository. The corpus and preprocessing pipeline are unknown, so the values above describe the released artifact but do not permit retraining it.

Dataset

Not documented. The training corpus, its size, sources, and any preprocessing or normalisation applied are not recorded in this repository.

Preprocessing pipeline

Not documented.

Usage

import fasttext
from huggingface_hub import hf_hub_download

path = hf_hub_download("Hailay/fasttext-tigrinya", "fasttext_tigrinya.bin")
model = fasttext.load_model(path)

vector = model.get_word_vector("ኣብ")          # 300-dimensional
neighbours = model.get_nearest_neighbors("ኣብ")

Installation

pip install fasttext huggingface_hub

Evaluation

Not documented. No evaluation methodology and no benchmark results are recorded for this model.

Intended use

Word and subword representations for Tigrinya NLP, including embedding initialisation for downstream models. Suitability for any specific application has not been evaluated.

Files

File	Description
fasttext_tigrinya.bin	FastText model binary

Citation

Not documented. No associated publication is recorded for this artifact.

License

Apache 2.0