MATHEMATICAL AND SOFTWARE SUPPORT FOR THE PROCESSES OF GENERATING AND PROCESSING UZBEK LANGUAGE WORD FORMS IN AUTOMATED TEXT PROCESSING SYSTEMS

Source abstract

This paper develops a corpus-evaluated framework for Uzbek nominal word-form generation and acquisition of stem classes from partially annotated examples. The model separates morphotactic order, stem alternation and surface realization. Instead of forcing an unsupported class decision, it retains all minimum-error hypotheses and returns a unique generated form only when those hypotheses agree. A minimax criterion identifies a diagnostic paradigm cell when further evidence is needed. An experiment on the frozen UD Uzbek-UT r2.15 resource uses a study-specific, genre-stratified sentence split and learns a lexicon of 873 lemmas from 1,474 training noun tokens. On 313 held-out noun tokens, compatible candidate recall reaches 58.15%, compared with 26.20% for surface memorization and 57.19% for a regular-class baseline. Recall is 94.30% on the 193 tokens whose lemmas occur in training. Acquired classes match the recall of unrestricted class enumeration while reducing disagreeing generation cells from 6,630 to 4,980. Error decomposition identifies unseen lemmas as the main coverage constraint. The study contributes an explicit treatment of incomplete morphological labels, a disagreement-aware generation interface and reproducible evidence separating lexical coverage from morphological realization. The results concern candidate recovery for nouns, not complete sentence-level grammatical analysis.