DEVELOPMENT OF A HYBRID ALGORITHM FOR DETECTING ADVERBIAL MODIFIERS IN UZBEK
Abstract
In this work, a hybrid algorithm and a software pipeline architecture were developed for automatically identifying adverbial modifiers in Uzbek texts. The solution consists of three stages: customized tokenization, transformer (BERT)-based part-of-speech (POS) tagging, and syntactic role extraction using a rule engine. To stabilize adverbial-modifier detection, a predicate (verb) identification module was introduced into the system as an anchor: the predicate is found using 6 formal rules, and adverbial modifiers are then labeled using 13 rules based on their linking conditions relative to the predicate. The system was evaluated on a test dataset and achieved 78% precision and an F1-score of 78% in adverbial-modifier detection. The results demonstrate the practical effectiveness of the hybrid (neural+rule-based) approach in low-resource settings, and emphasize that expanding rule coverage, enriching the dataset, and optimizing the module based on error analysis are the main directions for future work.