English-to-Bhojpuri Translator

Overview

This project presents a fine-tuned mBART model specifically designed for translating English text into Bhojpuri. Bhojpuri, as a low-resource language, has historically lacked large-scale machine translation resources. This work aims to bridge that gap by providing an end-to-end solution featuring a custom-built dataset and a fine-tuned multilingual model.

Access the deployed web application here:

• Live Demo

This project is inspired by previous work: ENG-TO-BHOJPURI GitHub Repository.

Image

Sample screenshot of the English-to-Bhojpuri Translator web application.

Key Components

• Model: BART-English-to-Bhojpuri-Alpha2
• Dataset: English-Bhojpuri Translation Dataset

Highlights

• First-of-its-kind Dataset: A curated English-Bhojpuri parallel corpus developed specifically to support machine translation and language modeling for Bhojpuri.
• Fine-tuned mBART Model: Utilizes the "facebook/mbart-large-50" architecture, adapted through focused fine-tuning on the custom dataset to ensure contextually relevant translations.
• Low-Resource Language Advancement: Contributes to Bhojpuri's digital presence by providing a strong foundational model for further research and applications.
• User-Friendly Interface: Built using Gradio Blocks with customized styling for a clean, intuitive user experience.

Technical Details

• Base Model: facebook/mbart-large-50
• Fine-tuning Library: Hugging Face Transformers
• Frontend Framework: Gradio
• Hardware Utilized: NVIDIA T4 GPU for model training

The model is configured with beam search (num_beams=5) to enhance the quality of generated translations. Maximum sequence length is set at 128 tokens to ensure efficiency without sacrificing translation quality.

Intended Applications

• Academic Research: Facilitates research in low-resource language processing and translation studies.
• Language Preservation: Supports initiatives aimed at preserving and promoting the Bhojpuri language in digital spaces.
• Content Localization: Can assist in adapting educational, cultural, or informational content into Bhojpuri.

Limitations and Future Work

• Early-stage Performance: As this is an Alpha release, the translations may sometimes exhibit literal interpretations or minor grammatical inaccuracies.
• Dataset Scope: Model performance is inherently tied to the diversity and size of the initial dataset. Expanding the dataset with more varied and context-rich data is a potential future direction.