fahim-ling/Sylheti-Bangla-English-Russian-German-Parallel-Corpus-for-NLP

Sylheti-Bangla-English-Russian-German Parallel Corpus for NLP

Welcome to the first open-source multilingual parallel corpus for the Sylheti (syl) language, engineered by a native Linguistics student. This dataset bridges Sylheti with four major global high-resource languages spanning three distinct language families (Indo-Aryan, Germanic, and Slavic) to support Computational Linguistics (CL), Natural Language Processing (NLP) research, and Large Language Model (LLM) fine-tuning.

📌 Project Overview

Sylheti is categorized as a low-resource language with a limited digital footprint. This project bridges the technological gap by building a highly structured multilingual parallel corpus. It is uniquely designed to support AI and LLM training, Machine Translation (MT) models, cross-lingual tokenizers, and dialectal analysis tools by mapping phonetic variations and traditional scripts across major global language families.

🚨 Ethical Plea & Strict Usage Notice

Aligning a severely low-resource language like Sylheti with four high-resource languages (Bangla, English, Russian, German) across different language families requires immense manual, syntactic, and computational effort.

Creating this sequence alignment—matching precise semantic equivalence, dialectal variations, and traditional Syloti Nagri (ꠍꠤꠟꠐꠤ ꠘꠣꠉꠞꠤ) script—is exceptionally complex and labor-intensive.

Please respect this grueling linguistic work:

• Do NOT scrape, republish, or commercialize this dataset without explicit attribution.
• Unauthorized text-scraping or utilizing this schema without citing the author is strictly prohibited under the CC BY-NC 4.0 license.

📂 Dataset Structure

The core dataset consists of parallel sentences aligned across multiple languages and is available in tab-separated format (.tsv):

• parallel_sentences.tsv : Sentence-level multilingual alignment for Machine Translation, multilingual NLP, and LLM fine-tuning.
• lexicons.tsv : Core vocabulary mapping, Sylheti variants, multilingual equivalents, linguistic metadata, and POS tagging.
• Pure Semantic Key (PSK) Indexing : The lexicons.tsv dataset uses the lXXXXX_vXX format, where lXXXXX represents the Lexical Concept / Synset ID and vXX represents the Sylheti Variant / Synonym ID. The parallel_sentences.tsv dataset uses the sXXXXX_vXX format, where sXXXXX represents the Sentence-level Semantic Equivalence Group ID and vXX represents the Sylheti Sentence Variant / Alternative Expression ID.
• Zero-Hallucination Policy: Missing or unverified linguistic data must be left completely empty (consecutive tabs \t\t). Literal strings like NaN, Null, or N/A and fabricated linguistic forms are strictly prohibited. This prevents tokenizer pollution and ensures clean auto-parsing for AI pipelines (Pandas/Hugging Face).

Technical Format Note: Although Hugging Face automatically categorizes text-based tabular data under the generic csv badge above, the actual production files in this repository are strictly formatted as TSV (Tab-Separated Values) to prevent commas in multilingual strings from corrupting the cell boundaries.

📊 Combined Data Preview (Lexicon & Sentence Structure)

Lexicon Schema Preview (11 Columns):

Index	Sylheti (Romanized)	Sylheti (Bangla Script)	Syloti Nagri (ꠍꠤꠟꠐꠤ ꠘꠣꠉꠞꠤ)	Bangla (Standard)	English (Standard)	Russian (Русский)	German (Deutsch)	Domain	Register	POS
l00001_v01	tai	তাই	ꠔꠣꠁ	সে (মহিলা)	She	Она	Sie	General	Informal	Pronoun

Parallel Sentences Schema Preview (10 Columns):

Index	Sylheti (Romanized)	Sylheti (Bangla Script)	Syloti Nagri (ꠍꠤꠟꠐꠤ ꠘꠣꠉꠞꠤ)	Bangla (Standard)	English (Standard)	Russian (Русский)	German (Deutsch)	Domain	Register
s00001_v01	tai oxon zaibogi	তাই অখন যাইবগি	ꠔꠣꠁ ꠅꠈꠘ ꠎꠣꠁꠛꠉꠤ	সে এখন চলে যাবে।	She will leave now.	Она сейчас уйдет.	Sie wird jetzt gehen.	Daily Life	Informal

🚀 Future Roadmap & Core Goals

• Scale to 10,000+ Parallel Sentences: Continually expand dataset rows to establish a world-class digital preservation corpus for AI.
• Granular Dialect Mapping: Systematically categorize sub-regional dialects (e.g., Sylhet Sadar, Sunamganj, Habiganj, Moulvibazar) using precise metadata tags.
• Phonetic & Acoustic Metadata (IPA): Develop a separate phonetic_metadata.tsv relational table mapped via strict decimal indices. This will integrate IPA transcriptions to preserve Sylheti tonogenesis & VOT, paving the way for advanced TTS & ASR technologies.
• Acoustic Audio Alignment: Collect high-fidelity spoken data for Automatic Speech Recognition (ASR) and Speech Synthesis (TTS).
• Nagri OCR Pipeline: Build character recognition models for ancient Nagri manuscripts (Pothi literature).

🤝 Contributing

This repository and corpus are maintained with strict access controls to preserve dataset integrity and prevent AI degradation.

• Strict Policy: External Pull Requests or dataset edits are accepted strictly by invitation only.
• To request research access or collaborate, contact the maintainer directly.

🧑‍💻 About the Author

I am Fahim Ahmed (cited academically as Ahmed Fahim), an undergraduate Linguistics student at Kursk State University, Russia, and a native Sylheti speaker. As an aspiring Computational Linguist, I am independently building this dataset to preserve, revitalize, and digitize the Sylheti language using modern AI pathways and LLM architectures.

📜 License

This dataset is strictly licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). You are free to share and adapt the material for non-commercial academic research only, provided you give appropriate credit.

📋 Citation

If you use this dataset, lexicon, or parallel corpus in your research, software, or AI models, please cite it exactly as follows:

@misc{fahim2026sylheti,
  author       = {Ahmed Fahim},
  title        = {Sylheti-Bangla-English-Russian-German Parallel Corpus for NLP},
  year         = {2026},
  version      = {1.0.0},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/fahim-ling/Sylheti-Bangla-English-Russian-German-Parallel-Corpus-for-NLP}
}