Developing A Context-Aware English-Ebira Translation System Using Fine-Tuned T5 Transformer and Position-Aware Idiom Detection

Abstract

Ebira is a language spoken by the Ebira people, known as "Anebira," a minority ethnic group spreading across five states and predominantly located in Kogi Central, Nigeria. The Ebira-speaking community of approximately 2.5 million faces a compounding linguistic crisis where about 35% of the older generation, though fluent in Ebira, remains largely excluded from the socioeconomic and knowledge advantages mediated through English, while younger generations are progressively abandoning their mother tongue in pursuit of English-driven social mobility, educational achievement, and economic inclusion, placing the language at risk of functional extinction without guaranteeing full English proficiency in return. Machine Translation (MT) had been adopted as a viable solution; however, the two existing English-Ebira rule-based systems suffer from poor fluency, limited accuracy, inability to translate idiomatic phrases, and significant loss of linguistic nuance. While NMT systems developed for related low-resource languages such as Nigerian Pidgin, and Igala show promising result with improvement in fluency and general translation accuracy, critical structural limitations persist: their models cannot distinguish literal from idiomatic content within the same sentence, fail on paraphrased idiomatic expressions, and produce translations that are structurally fluent but culturally hollow. To address these challenges, this research developed a unidirectional English-Ebira translation system built around a fine-tuned T5-base transformer model for general sentence translation with improved fluency, accuracy and proper handling of the nuances, incorporating a position-aware idiom detection layer for differential translation of mixed literal-idiomatic sentences, and a Sentence-BERT (SBERT) semantic embedding model with dictionary-enhanced detection for culturally validated idiomatic translation. The system was trained on 32,322 English–Ebira parallel sentence pairs, the T5-base model achieved a BLEU score of 40.95% with a 100% translation success rate. The hybrid system substantially outperformed both a rule-based baseline and the T5-only model on idiomatic content: on a 200-sentence idiomatic evaluation set, the T5-only model reached 8.5% of sentences at or above the 0.5 BLEU threshold, while the hybrid system achieved 90%, representing a relative gain of approximately 959%. On general translation, the hybrid system similarly outperformed the rule-based baseline, achieving 87% of test sentences at or above the 0.5 BLEU threshold compared to 81.5% for the baseline (McNemar chi-square = 6.37, p = 0.012). Human evaluation produced a mean cultural relevance score of 8.4/10, supporting hybrid symbolic-neural architectures for culturally appropriate low-resource translation.