A Novel Model of Deep Learning-Based Spelling Detection in Turkish

Abstract

The agglutinative language structure of Turkish makes it difficult for traditional rule-based spell checkers to detect errors, especially those that occur at the morphological level. In this study, DistillBERT-based model that can process context information bidirectionally is proposed to detect spelling errors in Turkish. A total of 313,644 incorrect-correct word pairs were generated by applying synthetic distortion techniques to words obtained from news texts, and these pairs were automatically labeled. Experimental results have achieved an accuracy rate of 95.40% in tests performed. The proposed method has been compared with Zemberek, Levenshtein-based methods, and BiLSTM revealing that the proposed model outperforms these traditional approaches. The results show that DistilBERT-based deep learning approaches supported by synthetic data generation provide an effective method for type detection. © 2025 IEEE.