TextCNN and Gated Recurrent Unit(GRU) with Additive Attention for Character-Level Yoruba Spell Correction

• Opeyemi O. ASAOLU

  Author

• Tomilayo F. ADEBIYI

  Author

• Kazeem M. OLAGUNJU

  Author

This Automatic spell correction for Yoruba, a tonal Niger-Congo language spoken by more than 40 million people, remains an open problem in low-resource natural language processing. This paper evaluates two deep learning architectures, a pure convolutional TextCNN and a Bidirectional GRU with additive attention (GRU+Attn) for character-level Yoruba orthographic error correction. Both models were trained on a 1,244,871-token corpus drawn from four open-access repositories: Masakhane NER, AfriSenti, MasakhaPOS, and MasakhaNEWS. A balanced dataset of 127,905 correction pairs was constructed by injecting five Yoruba-specific error types into a 1,000-word vocabulary. TextCNN reached 83.9% test accuracy in 1.82 minutes of Graphics Processing Unit (GPU) training. GRU+Attn reached 83.6% in 2.49 minutes. Both architectures corrected character insertion errors above 93% accuracy. Tone mark omission was the hardest category for both, peaking at 77.3% (TextCNN), a result consistent with the information-theoretic difficulty of resolving homographs without sentential context. These results extend the existing Yoruba benchmark and establish TextCNN as a deployment-practical alternative to stacked recurrent models.

• 22_Asaolu et al. (2026)

Copyright (c) 2026 Opeyemi O. ASAOLU, Tomilayo F. ADEBIYI, Kazeem M. OLAGUNJU (Author)

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.