Improving Machine Translation Capabilities by Fine-Tuning Large Language Models and Prompt Engineering with Domain-Specific Data

This study examines the applicability and performance of large language models (LLMs) in the field of machine translation for in-domain texts, with a particular focus on tine-tuning LLMs, few-shot prompting, and word vector based example sentence search methods. The aim of the study is to determine the extent to which the few-shot technique can improve translation quality for domain-specific texts. Our results indicate that the few-shot learning approach consistently improved translation quality across all examined LLM systems, with performance enhancements ranging from 10% to 25% in BLEU scores. Surprisingly, the word vector-based method, which uses the vectorial representation of words to select translation examples, did not perform as well as the character similarity based fuzzy matching technique. The study discusses the performance of various systems, highlighting significant advancements achieved through tine-tuning and few-shot prompting.