Detection and correction of spelling errors in Uzbek texts based on machine learning algorithms

Abstract

This study addresses the problem of detecting and correcting spelling errors in Uzbek texts. Due to the complex morphological structure and agglutinative nature of the Uzbek language, traditional spell-checking methods do not provide sufficient accuracy. Therefore, this research employs the Levenshtein distance algorithm to measure word similarity and utilizes neural network-based language models for contextual correction. KenLM (a statistical language model), LSTM (Long Short-Term Memory), and BiLSTM (Bidirectional LSTM) approaches were used as language models. A text corpus of 80 million words was collected and analyzed for model training. The test results indicate that the BiLSTM model achieved the highest accuracy (90.09%) in correcting spelling errors, while the LSTM model recorded 84.62% accuracy. The KenLM model demonstrated an accuracy of 62.21% as well. These findings highlight that deep learning models capable of contextual analysis can significantly improve the automatic detection and correction of spelling errors in the Uzbek language. Based on the study results, future research plans include the application of transformer models, the expansion of annotated corpora, and the development of models that consider various morphological characteristics of the Uzbek language.