HYBRID DEEP MODEL FOR UZBEK LANGUAGE PUNCTUATION PREDICTION

Automatic punctuation restoration is one of the important tasks in natural language processing, especially for low-resource languages such as Uzbek, where this problem remains particularly relevant. The lack of sufficiently annotated corpora negatively affects the performance of downstream applications, including speech recognition, machine translation, and semantic text analysis. This study proposes a hybrid deep learning model designed to predict punctuation marks in Uzbek texts. The proposed approach combines BERT-based contextual vector representations, BiLSTM for modeling sequential dependencies, and a rule-based post-processing stage grounded in linguistic knowledge. This architecture effectively leverages the semantic capabilities of transformer models and the temporal dependency modeling strengths of recurrent networks, while the rule-based correction component improves the accuracy of punctuation detection in ambiguous boundary cases. Experimental results obtained on an annotated Uzbek corpus demonstrate that the proposed model outperforms existing neural and statistical approaches in terms of precision, recall, and F1-score. The findings confirm that integrating deep neural architectures with linguistic rules significantly enhances punctuation restoration performance for low-resource languages. This work presents a practical and extensible approach for advancing Uzbek language processing and improving the accuracy of various applied NLP systems.