BRIDGING GAPS IN LOW-RESOURCE LANGUAGES: A MACHINE LEARNING APPROACH TO PRONOMINAL ANAPHORA RESOLUTION

This study proposes the Kazakh Coreference Adaptation (KCA) model, a hybrid framework for resolving pronominal anaphora in Kazakh, a morphologically rich and low-resource language for which existing coreference systems and large language models (LLMs) remain unreliable. The study aims to evaluate whether integrating explicit linguistic constraints with supervised machine learning can outperform large language models in this setting. The novelty of the approach lies in combining deterministic morphological filtering with probabilistic ranking trained on a newly annotated subset of the Kazakh National Corpus. The model integrates rule-based constraints-covering case, number, person agreement, and syntactic accessibility-with a RandomForest classifier trained on 4,200 sentences across five genres. The linguistic stage eliminates grammatically implausible candidates, while the machine-learning component ranks the remaining antecedents using morphological, syntactic, discourse, and semantic features. Experimental results show that KCA achieves 79% accuracy on long-context anaphora cases, exceeding GPT-4o (70%) and o1-mini (73%). The hybrid architecture demonstrates clear advantages in resolving long-distance and morphologically complex structures. These findings highlight the importance of morphology-aware modeling in agglutinative languages and provide a reproducible dataset and methodological framework applicable to other low-resource settings.