如何改進英語至 Kinyarwanda 翻譯項目

During my NLP studies, I worked on a project to translate English into Kinyarwanda using GPT-2. I used a dataset from Hugging Face with around 3,000 English–Kinyarwanda sentence pairs.

However, GPT-2 was giving me wrong translations, and sometimes the sentences did not make sense. After learning more about Large Language Models (LLMs), I wanted to find a model that could work better for my project.

What I Learned About LLMs Today;

Today, I learned that LLMs learn patterns in language by predicting the next token. We also learned about tokenization, embeddings, Transformers, and how models generate text.

This helped me understand why GPT-2 can generate text but may not always translate correctly. GPT-2 is mainly designed to predict the next token, while some models are specifically designed for translation. I realized that choosing the right model is important for getting better results.

Models That Could Improve our Project;

NLLB-200

NLLB-200 is a multilingual translation model developed by Meta AI. It supports Kinyarwanda and is designed to translate between different languages. I think it could be a good choice for my project because translation is its main purpose.

M2M100

M2M100 is another multilingual model that translates directly between different languages. I could test it with my dataset and compare its translations with GPT-2.

mBART

mBART is a multilingual Transformer model that can be used for translation. I could explore fine-tuning it with my English–Kinyarwanda sentence pairs to see if it improves the results.

Which Model Would I Choose?

I would choose NLLB-200 because it is designed for multilingual translation and supports Kinyarwanda. I also found a version fine-tuned for English-to-Kinyarwanda translation by MbazaNLP.

Instead of just choosing a model because it sounds good, I would test it using the same sentences as GPT-2 and compare the results. This would help me see which model gives more accurate and meaningful Kinyarwanda translations.

This lesson helped me understand that not every language model is the best choice for every task. GPT-2 helped me start my project, but now I want to try NLLB-200 to see if I can improve my translations. I learned that understanding how models work can help us choose better tools for our projects.