English innovativejournal825@gmail.com
JAIMT Logo

JAIMT

Journal of Artificial Intelligence and Modern Technology (JAIMT)

Published article details, abstract, issue information, DOI, and downloadable manuscript file.

E-ISSN: 1595-6281 Bimonthly Publication Submit: innovativejournal825@gmail.com
Publication Details

DEVELOPMENT OF AN ENHANCED SPELLING ERROR DETECTION AND CORRECTION SYSTEM USING NATURAL LANGUAGE PROCESSING TECHNIQUE

Author(s)
Article Type Research Article
Pages 48-61
Issue Vol.8 No.1, 2026
Publication Date

Abstract

Spelling is an important part of human communication, as it plays a key role in conveying meaning clearly and accurately in written language. Errors in spelling can lead to misunderstanding, reduce the quality and credibility of a text. Most existing spelling correction systems rely on dictionary-based and rule-based approaches, which are limited in understanding contextual meaning and often fail to detect real-word errors where words are correctly spelled but used inappropriately. Additionally, such systems lack personalization and do not adapt to individual writing styles. This study presents the development of an enhanced spelling error detection and correction system using natural language processing techniques. The proposed system adopts a context-aware and personalized approach based on Bidirectional Encoder Representations from Transformers (BERT), enabling it to understand the contextual meaning of words within sentences. The system was engineered as a modular pipeline comprising text preprocessing (normalisation and WordPiece tokenisation), a dual detection layer for out-of-vocabulary and contextual errors, a correction layer that generates and ranks candidate replacements using maskedlanguage-model probabilities together with edit-distance and phonetic similarity signals, and a feedback loop through which user decisions are captured and fed back into model refinement. The system is implemented using a modular architecture consisting of a user interface, application layer, and model inference layer. A structured dataset of Misspelled Dataset from Kaggle of 7840 correct words, misspellings of 37105 words, and contextual sentences was pre-processed and split into training and testing sets in an 80:20 ratio. The model was evaluated using precision, recall, and F1-score. Results show that the system achieved a precision of 90.0%, recall of 89.8%, and F1-score of 89.7%, the ROC (Receiver Operating Characteristic) computed using a one vs rest strategy in spelling, duplicate, capitalization, and context error-type class achieved an Area Under curve values all above 0.9, demonstrating excellent discriminative power and improved performance. An entropy-based active learning strategy was additionally shown to reach 95.8% labelling efficiency with only a third of the queries required by random sampling, underscoring the practicality of continuous, low-cost model improvement. The study concludes that integrating contextual deep learning with personalized learning enhances spelling correction systems