Tajik FastText Word Embedding Model
This repository contains a pretrained FastText model for the Tajik language, trained on a large corpus of Tajik texts. The model supports subword information, allowing it to generate embeddings even for rare or unseen (OOV) words.
The model is suitable for use in various NLP tasks such as:
- Semantic analysis
- Text classification
- Machine translation
- Synonym detection and thesaurus building
- Enhancing other models through embedding initialization
Licensed under the MIT License, which allows free usage in both research and commercial applications.
📊 Model Overview
| Parameter | Value |
|---|---|
| Model Type | FastText (with subwords) |
| Vector Size | 300 |
| Vocabulary Size | 145,232 |
| OOV Support | Yes |
| Context Window | 5 |
| Min Word Count | ≥ 5 |
📚 Training Corpus
Books (Total: 99)
- Programming: 6
- History: 4
- Religion: 12
- Scientific: 3
- Children's literature: 6
- Prose: 19
- Poetry: 21
- Textbooks: 28
Articles (Total: 134,497)
- Asia-Plus: 20,471
- Khovar: 21,557
- Ovozi Tojik: 7,495
- Farazh: 4,679
- Wikipedia: 80,295
Total Corpus Statistics
- Documents: 134,596
- Tokens: 33,535,383
- Unique Lemmas: 649,308
🧪 Model Comparison with Meta FastText
We evaluated our model against Meta’s pretrained FastText using semantic similarity and Spearman correlation:
| Model | Spearman Correlation | OOV Support |
|---|---|---|
| FastText (Meta) | 0.703 | Yes |
| FastText (ours) | 0.622 | Yes |
While Meta FastText achieves better overall performance, our model demonstrates strong results on Tajik-specific morphology and semantics.
🔍 Example Similar Words
| Word | Nearest Neighbors (FastText) |
|---|---|
| кӯдак | кӯдаку(0.82), хурдкӯдак(0.81), кӯдакам(0.81), кӯдакат(0.81), кӯдаке(0.81) |
| муаллим | муаллиме(0.90), муаллимат(0.89), муаллимин(0.89), муаллиму(0.88), муаллима(0.88) |
| об | оби(0.79), обро(0.74), обмӯрии(0.70), обшустаи(0.68), обшуста(0.66) |
| мард | марда(0.87), мардхӯ(0.85), мардвор(0.85), мардро(0.83), зан(0.82) |
| деҳа | деҳайи(0.83), деҳаю(0.80), деҳавз(0.78), деҳакӣ(0.76), деҳодеҳ(0.74) |
| китоб | китобӣ(0.84), китобгуна(0.83), китобча(0.81), китобсӯзӣ(0.81), китобро(0.81) |
| меҳмон | меҳмонӣ(0.86), меҳмоншо(0.85), меҳмонат(0.83), меҳмонҳона(0.82), меҳмони(0.82) |
| шаҳр | шаҳрӯ(0.82), шаҳрча(0.80), бушаҳр(0.79), шаҳрат(0.79), навшаҳр(0.79) |
| падар | падаршӯ(0.89), падарӣ(0.84), падаршӯву(0.84), падаре(0.84), падаршон(0.83) |
| модар | модаршӯ(0.86), модаршӯяш(0.83), модару(0.81), модаре(0.81), модарвор(0.80) |
🧩 Handling OOV (Out-of-Vocabulary) Words
FastText supports generating vectors for unknown words via subword units (n-grams). Here are some examples:
| Unknown Word | Closest Matches (FastText) |
|---|---|
| кӯдакона | кӯдаконаи(0.82), кӯдаконат(0.81), кӯдаконае(0.81) |
| меҳмонамон | меҳмон(0.77), меҳмонҳо(0.77), меҳмонам(0.76) |
| муаллимон | муаллимони(0.89), муаллимоне(0.88), муаллимону(0.83) |
| деҳоти | дарҷамоати(0.79), чамоати(0.74), ҷамоати(0.81) |
| саводнок | саводнокӣ(0.88), саводнокиву(0.85), саводнокии(0.84) |
📌 Features for Tajik Language
Our model performs well on:
- Semantic similarity: e.g., "мард" ↔ "зан", "китоб" ↔ "китобгуна"
- Morphological variants: e.g., "кӯдак" → "кӯдаку", "кӯдаки"
- Rare/compound words: thanks to subword representations like "саводнок", "деҳоти"
💡 Usage Example
from gensim.models import FastText
model = FastText.load("tajik_fasttext.model")
vector = model.wv["падар"] # Get vector for a word
similar_words = model.wv.most_similar("модар") # Find similar words
🗂️ Files Included
| File | Description |
|---|---|
tajik_fasttext.model |
Gensim FastText model file |
*.npy files |
Supporting NumPy arrays for vectors |
📚 Citation
If you use this model, please cite:
@misc{ArabovMK_Tajik_FastText,
author = {ArabovMK},
title = {Tajik FastText Word Embeddings},
year = 2025,
publisher = {Hugging Face},
url = {https://huggingface.co/ArabovMK/tajik-fasttext-model}
}
Last updated: 2025-05-10