Enhancing Yoruba Text Autocompletion with an Attention-Augmented Recurrent Neural Network

Abstract

The digital divide in Natural Language Processing (NLP) is particularly pronounced for low-resource, morphologically complex languages like Yoruba. This paper addresses the challenge of developing an effective text autocompletion system for Yoruba, a language characterized by its tonal diacritics and agglutinative structure, which are poorly handled by conventional models. A character-level Recurrent Neural Network (RNN) architecture enhanced with a multi-head attention mechanism to overcome the limitations of standard RNNs in capturing long-range contextual dependencies was proposed. A curated dataset of 4,431 Yoruba words was used for training and evaluation. The proposed RNN + Attention model was rigorously evaluated against a baseline RNN, demonstrating a significant 82.5% improvement in model confidence, achieving a perplexity of 2.21 compared to the baseline's 12.67. The model also achieved perfect Top-K accuracy and Mean Reciprocal Rank, indicating its high precision in ranking correct suggestions. The results conclusively show that integrating an attention mechanism is a pivotal architectural enhancement for sequence prediction tasks in Yoruba, leading to a robust and contextually aware autocompletion system. This work provides a validated framework for building efficient NLP tools for low-resource languages.