Hybrid Transformer and Linguistic Knowledge-Based Morphological Analysis for Uzbek Texts
Abstract
Morphological analysis is a foundational task for Uzbek natural language processing because Uzbek is a morphologically rich agglutinative Turkic language in which a single word may contain a stem and several derivational or inflectional suffixes. This paper presents a hybrid Transformer and linguistic knowledge-based morphological analysis approach for Uzbek texts. The proposed system combines a ByT5-based sequence-to-sequence morphological segmentation model with a symbolic linguistic knowledge base that contains allomorph mappings, suffix family priorities, contextual compatibility rules, and lemma formation constraints. The Transformer component predicts morpheme boundaries for Latin and Cyrillic Uzbek word forms, while the linguistic component normalizes the predicted morphemes, assigns candidate morphological tags, performs contextual disambiguation, and constructs lemma candidates. The method is evaluated on Uzbek morphological segmentation data using Exact Match, Token Accuracy, and Weighted F1. Experimental results show high segmentation performance for Latin-script texts and strong performance for Cyrillic-script texts; the integrated hybrid morphological analyzer currently achieves approximately 96% accuracy, while the symbolic interpretation layer improves grammatical consistency and interpretability of the final analysis. The proposed neuro-symbolic architecture provides a practical and extensible foundation for Uzbek morphological analysis and downstream linguistic processing.