Neural Machine Translation of Sinhala to English

Abstract

This addresses the challenges of translating Sinhala in the form of Singlish, a unique and informal variation of Sinhala written in English letters, into standard English, aiming to bridge the communication gap for Sinhala-speaking individuals. The research explores the application of advanced transformer architectures such as T5, mBART, and MarianMT, each chosen for their proven capabilities in handling multilingual and low-resource language tasks. The study involves training and fine-tuning these models on a curated Singlish and English parallel sentences including dataset and evaluates their performance across several key metrics, including Quantitative and Qualitative analysis. The system developed as part of this research integrates practical features like speech-to-text for Singlish input, text-to-speech for English output, and translation history for user convenience, making it a comprehensive tool for real-time translation needs. The thesis not only aims to produce an effective translation system but also contributes valuable insights into optimizing transformer models for low-resource languages, offering benchmarks and techniques for future research in neural machine translation. By combining cutting-edge technology with a user-centric design, this research seeks to enhance accessibility for Sinhala speakers and sets the stage for advancing machine translation for underrepresented languages and dialects.