Belarusian translation improves with better data cleaning and fine-tuning
ReMova: Fine-tuning LLMs for English to Belarusian translation
Computation and Language
Summary
Translating English to Belarusian faces challenges because of two different writing styles and common spelling mistakes online. The authors created a special way to clean up translation training data that fixes these problems. They found that cleaning the data helps computers learn to translate better, especially for large language models. This shows that cleaner data is very important for good Belarusian translation.
What this means in practice
- •For machine translation engineers: Improve English to Belarusian translation systems by using a specialized data cleaning pipeline to reduce errors from orthographic differences and misspellings.
- •For localization service providers: Enhance translation quality for Belarusian market content by integrating fine-tuned large language models relying on cleaner training data.$Commercial implications: Enables companies to offer higher-quality Belarusian translations in localization services, improving customer satisfaction and market reach.
Authors
Mikita Pilinka, Aliaksandr Kliujeŭ, David Samuel, Yves Scherrer
Abstract
This paper presents a Belarusian-specific data-cleaning pipeline and fine-tuning for English-Belarusian machine translation. Our cleaning pipeline distinguishes itself from others by employing a correction tool that addresses the issue of the two orthographies of the Belarusian language, noise in the training data, interference from other languages and other misspelling issues common in Belarusian on the internet. A matched ablation on unfiltered training data shows substantial benefits from filtering for all fine-tuned models, with the LLM-based models gaining roughly twice as much from filtering as the dedicated encoder-decoder MT system, supporting the view that for Belarusian MT one of the primary bottlenecks is data quality.