Introduction

I am excited to share my latest project: a bidirectional neural machine translation model for Twi (a language spoken in Ghana) and English. This model was fine-tuned from Helsinki-NLP's OPUS-MT multilingual base model and is now available on Hugging Face.

Why Twi?

Twi is one of the most widely spoken languages in Ghana, with millions of native speakers. However, like many African languages, it remains under-resourced in the NLP community. This project aims to bridge that gap by providing accessible translation tools for Twi speakers.

Model Architecture

The model is based on the MarianMT architecture, specifically fine-tuned from Helsinki-NLP/opus-mt-mul-en. Key specifications:

Parameters: ~74M
Format: Safetensors
License: CC-BY-NC-4.0

Training Details

I trained this model on the Ghana-NLP/TWI_ENGLISH_PARALLEL_TEXT dataset, which contains 10,370 professionally translated sentence pairs covering domains like civic life, health, agriculture, education, and culture.

Training Configuration

Epochs: 5
Batch Size: 16
Learning Rate: 5e-5
Hardware: NVIDIA T4 GPU (Google Colab)
Training Time: ~25.6 minutes

Performance Results

Twi → English Translation

BLEU Score: 10.50
chrF Score: 32.30

English → Twi Translation

BLEU Score: 6.32
chrF Score: 28.74

While these scores are lower than high-resource language pairs, they represent a significant achievement for low-resource African language NLP.

Usage

from transformers import MarianMTModel, MarianTokenizer

model = MarianMTModel.from_pretrained("ninte/twi-en-marianmt")
tokenizer = MarianTokenizer.from_pretrained("ninte/twi-en-marianmt")

text = "Akwaaba! Wo ho te sɛn?"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, num_beams=4, max_length=128, no_repeat_ngram_size=3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Future Work

I plan to continue improving this model by:

Fine-tuning on NLLB-200 for better native Twi support
Training separate direction-specific models
Using back-translation for data augmentation

Try It Out

The model is available on Hugging Face: ninte/twi-en-marianmt

This work was made possible thanks to GhanaNLP's dataset and Google's computational resources. I'm excited to contribute to the growing ecosystem of African language technology!

Kwabena Obeng

Comments (0)

Leave a comment