Overview
This Hindi-to-Manipuri (Bengali script) translation model is built on the foundation of the facebook/mbart-large-50-many-to-many-mmt model, a powerful multilingual machine translation model. It has been fine-tuned specifically to handle translation tasks between Hindi and Manipuri, with a focus on generating accurate and contextually appropriate translations in the Bengali script of Manipuri. This model aims to serve users who require reliable and efficient translations between these languages, making it especially valuable for linguistic research, educational content, and cross-cultural communication.
Model Summary
- Model Type: Text2Text Generation
- Objective: Machine Translation
- License: mit
- Finetuned from: mBART-large-50.
mbart-large-50-many-to-many-mmt
- Training Dataset: abhishek2475/hi-mni-dataset
- Developer: Abhishek Aryan
Application and Use Cases
- Text Translation: easily translate from Hindi to Manipuri(Bengali)
- Corpus Mangaement : for managing parallel corpus between Hindi and Manipuri
How to use the Model:
Using pipeline
1from transformers import pipeline
2
3pipe = pipeline("text2text-generation", model="abhishek2475/mb-Hindi-to-manipuri")
Load Model Directly
1from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
2
3tokenizer = AutoTokenizer.from_pretrained("abhishek2475/mb-Hindi-to-manipuri")
4model = AutoModelForSeq2SeqLM.from_pretrained("abhishek2475/mb-Hindi-to-manipuri")
Dataset Information
The model was finetuned on the abhishek2475/hi-mni-dataset , which contains :
- Parallel Corpus: between Manipuri(bengali) and hindi Language.
- Size: 21400 parallel manually translated lines
- Extracted: using newspaper and texts
Model Architecture
- Transformer Encoded Layers: Utilizes self-attention mechanisms.
- Positional Encodings: Helps the model understand the corpus.
Export a file
You can export the current file by clicking Export to disk in the menu. You can choose to export the file as plain Markdown, as HTML using a Handlebars template or as a PDF.
Training Details
- Learning Rate: 0.0000001
- Batch Size: 64
- Epochs: 100
- Training Time: 1 hr 36 min
Evaluation
The model was evaluated using the BLEU score, with the following :
BLEU Score: 9.10
Limited by hardware resources, training set size, and vocabulary definition, leading to a relatively low BLEU score.
System Workflow
The whole system was implemented on a Corpus Management Tool
- Frontend: HTML CSS
- Backend: Python Django
- Database: SQL Lite
- Model: Deployed via Pytorch and TensorFlow frameworks
Strengths and Limitation
Strengths
- Local translation system : Allows state of the art translations for locals as there is scarcity of resources.
- Pretrained on mBART: Allows for efficient transfer learning and robust generalization.
Limitations
- ** Low resources**: Since Manipuri is a low resource language the dataset is pretty small which effects the model performance vastly.
- Data Bias: Like all machine learning models, its predictions may reflect biases present in the training data.
Conclusion and Future Work
The Hindi-to-Manipuri (Bengali script) translation model offers a significant step forward in addressing the linguistic gap between these two languages. By fine-tuning the facebook/mbart-large-50-many-to-many-mmt model, this project demonstrates how advanced multilingual models can be adapted to perform specialized translation tasks with improved accuracy and fluency. The model has the potential to support applications in education, digital content creation, and cross-linguistic communication, contributing to the preservation and accessibility of Manipuri language content in the Bengali script.
Future Work
-
Dataset Expansion: Including larger and more diverse datasets from various domains such as literature, news, and social media could improve the model's ability to handle complex and domain-specific translations.
-
Support for Additional Scripts: While this model focuses on the Bengali script for Manipuri, adding support for other scripts like Meitei Mayek can broaden its usability and impact.
-
Fine-Tuning on More Language Pairs: Extending the model's capabilities to handle translations between Hindi and other regional languages, or between Manipuri and other Indian languages, could make it a more versatile tool.
