Multi-Level Model for Solving Reference Relations in the Kazakh Language

This article addresses the lack of robust tools for resolving reference relations in the Kazakh language, a low-resource language in the field of Natural Language Processing (NLP). To solve this issue, we propose a multi-level model based on domain-specific knowledge. The system automatically establishes connections between referents and anaphors by integrating morphological, syntactic, and semantic levels of analysis. The novelty of the model lies in its integration of an ontological knowledge base and a custom-built semantic dictionary, which enable accurate referential identity recognition even in complex sentence structures. The proposed architecture comprises three main modules: morphological analysis, candidate antecedent selection, and semantic similarity evaluation. The system was implemented using Python and C++ and tested on real-world Kazakh-language data derived from scientific and technical texts. As a result of the experiments, the model achieved an F1-score of 0.86, demonstrating high accuracy in identifying coreferential links. This performance metric indicates the model’s practical effectiveness in handling reference resolution tasks. This work represents one of the first comprehensive systems dedicated to referential resolution in the Kazakh language and may serve as a foundation for future multilingual intelligent NLP applications.