DEVELOPMENT OF A BILINGUAL SPELL CHECKER SYSTEM USING TRANSFORMER-BASED MODELS

ABSTRACT Language is a uniquely human tool shaped by informal rules, which can be difficult for computers to interpret especially in spellchecking. While essential in NLP and search tools, many spellcheckers today either lack accuracy or speed. This project tackled those challenges by developing a bilingual spellchecker for English and German using transformer-based models.The goal is to build a system that can identify and correct spelling errors in English and German. Key objectives include: creating a user-friendly interface using Django, JavaScript, and TailwindCSS, implementing transformer-based spellchecking with Python and evaluating the system using standard performance metrics. Transformer models (“oliverghur/spelling-correction-english-base” and “german-base”) were used. English data was fine-tuned on 10,000 sentences; German was tested on 50. Preprocessing involved cleaning, tokenizing, normalizing, and adjusting sequence length. The backend was built with Django, and the frontend with TailwindCSS and JavaScript.The English model achieved 81.15% accuracy and a 99.76% AUC ROC score. The German model showed 90.84% recall, 83.51% F1 score, and 60.00% accuracy. The interface lets users input text and get corrections in the selected language. The preprocessing pipeline played a key role in the model’s performance. This project presents a scalable solution for bilingual spellchecking. While English performed well, the German model shows the need for better datasets or bilingual training. Future improvements could include user accounts, feedback features, and building more robust models from scratch.