"Language verY Rare for All"

Translation for rare languages now achievable with minimal computing power.

LYRA combines transfer learning, data standardization, and retrieval augmentation to enable high-quality machine translation for rare languages using a single GPU.

https://arxiv.org/abs/2412.13924

## Original Problem 🤔:

→ Rare languages lack sufficient data for training neural machine translation models

→ Limited computational resources make it challenging to develop translation systems for low-resource languages

→ Existing translation tools don't support many rare languages like Monégasque

-----

## Solution in this Paper 🛠️:

→ LYRA introduces a three-pronged approach using a single GPU setup

→ Transfer learning leverages grammatical similarities between related languages (French-Italian to French-Monégasque)

→ Data standardization fixes inconsistencies in capitalization, punctuation, and quotation marks

→ Retrieval Augmented Generation finds similar sentence pairs from training data to improve translation quality

→ Implementation uses three model variants: LYRA-L (Llama-3.1-8B), LYRA-G (gemma-2-9b), and LYRA-M (Mistral-Nemo-Instruct)

-----

## Key Insights 💡:

→ Data quality improvements significantly boost translation performance

→ RAG enhances translation quality towards French across all models

→ Transfer learning benefits depend on language relationships

→ Single GPU training makes rare language translation more accessible

-----

## Results 📊:

→ LYRA-G with RAG achieved highest BLEU scores (58.10) for Monégasque to French translation

→ Data standardization improved performance across all models

→ NLLB-200 1.3B and LYRA-G showed comparable performance for French to Monégasque translation