NEURAL NETWORK APPROACHES FOR ACCURATE AFAN OROMO SPELL CHECKING AND CORRECTION

Abstract

Afan Oromo, a widely spoken Cushitic language, lacks advanced natural language processing (NLP) tools like spell checkers due to limited resources and linguistic expertise. Both native and non-native speakers face challenges in writing Afan Oromo correctly, partly because its Latin-based Qubee script was adopted in 1991. Traditional spell-checking methods, such as dictionary lookup and rule-based approaches, are inadequate for Afan Oromo’s highly inflectional morphology. This thesis proposes a neural network-based spell checker using a sequence-to-sequence (Seq2Seq) model with Long Short-Term Memory (LSTM) layers. A corpus of 596,948 words was collected from BBC Afan Oromoo using Sketch Engine, ensuring compliance with BBC’s terms of service. The model was trained to detect and correct spelling errors, achieving 100% error recall and 52.47% precision. This work is the first to apply neural networks to Afan Oromo spell checking, offering a scalable solution for under-resourced languages.