Muluken Hussen Asebel, Shimelis Getu Assefa and Mesfin Abebe Haile
Machine translation has made significant progress in automating the conversion of human languages via computational methods. However, achieving human-level performance remains challenging, particularly for languages such as Amharic. This paper aims to bridge this gap by integrating prior knowledge, particularly the syntactic structure of the source language, into graph neural networks for English-to-Amharic machine translation. Our objective is to systematically evaluate the effectiveness of integrating source language syntactic information into GNNs to improve English to Amharic machine translation quality. We conduct a thorough review of the relevant literature and describe the preprocessing steps for both existing and newly collected parallel corpora used in training. Our approach involves preprocessing data and discussing the proposed Graph2Seq models. The experimental results demonstrate a notable 4.56% increase in the bilingual evaluation understudy (BLEU) score compared with the baseline score, indicating a significant improvement in translation quality. Moreover, our models exhibit a 1.98% enhancement in the BLEU score over previous attempts, highlighting the value of integrating syntactic information into graph neural networks. Through meticulous experimentation and analysis, we illustrate the efficacy of incorporating source language syntax into GNNs for enhancing English-to-Amharic machine translation. This study advances machine translation systems, particularly for low-resource languages, and lays the foundation for future research in integrating syntactic knowledge across diverse linguistic tasks and languages.