A Transformer-Based Framework forDomain-Sensitive Amharic to English MachineTranslation with Character-Aware SubwordEncoding

This paper proposes a domain-adapted neural machine translation (NMT) system for Amharic-to-English translation, focusing on the issues of low-resource translation in a richly morphologically inflected language. We focus on the religious domain with the Tanzil corpus, a structured collection of Quranic verses which are translated into Amharic and English for coherence and semantic correspondence. To address the shortcomings of the traditional word-level tokenization of Amharic, we implement character-level subword tokenization using the SentencePiece model, which is better suited for rare and compound words. Our model harnesses a Transformer based encoder-decoder model together with multi-head attention and feedforward layers induction over the parallel corpus of poems in English and Amharic.The model achieved 59.03 BLEU score on the test set, greatly exceeding the classical RNN+Attention baselines which have been shown to have poor performance in low-resource settings. The strong score illustrates that an effectively tuned baseline Transformer model, in combination with domain-specific corpora and sophisticated subword methods, can perform well in translation tasks for under-resourced languages. The research provides a foundational, reproducible, and scalable framework that is linguistically-informed for Amharic-English translation, and it can be extended in the future, with additional extensions to other Semitic and morphologically rich languages. Our results highlight the value of domain adaptation and subword-aware architectures in advancing NMT for low-resource language communities.