Comparative Analysis of Neural Sequence Labeling Approaches for Uzbek Morphological Modeling

Abstract

Uzbek is an agglutinative language where a single root can take multiple suffixes, creating millions of word combinations. This complexity makes text processing difficult for standard NLP algorithms and older rule-based systems. To solve this problem, our study evaluates three deep learning models, BiLSTM, BiGRU, and Transformer approaching morphological analysis as a character-level sequence labeling task. We trained and tested these models on a new dataset of 80 K real Uzbek words. Each word is segmented using the Complete Set of Endings (CSE) rules, allowing us to accurately tag 16 different morphological features. Experimental results indicate strong performance across all three architectures, with stemming accuracy consistently around 95%. Although the BiGRU model is slower than the other architectures in terms of latency, it remains the most suitable choice for practical use because it achieves the highest accuracy and F1-score. Additionally, the Transformer effectively handles exceptionally long words with many suffixes. Overall, this work provides a fast, accurate, and practical neural network framework for processing the Uzbek language.

More by Ulugbek Salaev

• A SINGLE-LAYER FEEDFORWARD NEURAL NETWORK MODEL FOR MORPHOLOGICAL STEMMING OF UZBEK WORDSIlmiy axborotnoma/Scientific Journal of Samarkand University · 2026
• Neural Sequence Models for Uzbek Morphological StemmingInternational Scientific-Technical Conference on Actual Problems of Electronic Instrument Engineering · 2025 · Citations: 7
• A Morphology-Aware Neural Network Model for Uzbek Text ClassificationInternational Scientific-Technical Conference on Actual Problems of Electronic Instrument Engineering · 2025 · Citations: 1
• A Feedforward Neural Model for Morphological Stemming of Uzbek WordsInternational Conference on Computer Science and Engineering · 2025 · Citations: 3