FORMAL AND NEURAL APPROACHES TO MORPHOLOGICAL ANALYSIS IN AGGLUTINATIVE LANGUAGES: EVIDENCE FROM KAZAKH

Abstract

Automatic morphological analysis remains a challenging task for agglutinative languages because of their rich inflectional systems, productive derivation, and complex morphophonological rules. Recent neural models, especially transformer-based architectures, have demonstrated impressive empirical performance. However, they frequently exhibit a deficiency in linguistic transparency and encounter challenges in systematic generalization inside low-resource environments. This research offers a comparative and integrative examination of formal (rule-based and finite-state) and neural (KazBERT-based) methodologies for morphological analysis, utilizing the Kazakh language as a case study of low-resource agglutinative morphology. Initially present a formal morphological model that distinctly represents root-affix structure, vowel harmony, and morphotactic restrictions. We next test many neural architectures for morphological disambiguation and tagging, such as KazBERT coupled with CRF-based decoding. In addition to typical accuracy measurements, we do a comprehensive error taxonomy and linguistic analysis, investigating how various model classes manage ambiguity, infrequent forms, and extended affix chains. The findings indicate that whereas neural models excel in surface-level accuracy compared to exclusively rule-based systems, they demonstrate consistent deficiencies in morphologically intricate and infrequent constructs. On the other hand, formal models show better generalization based on language limitations. Based on these results, we suggest a hybrid morphology-aware framework that adds symbolic restrictions to neural inference. This framework consistently improves results in a variety of assessment contexts. The study demonstrates that effective morphological analysis of agglutinative languages requires the integration of neural representation learning with explicit linguistic structure. The results are not tied to any one language and have wider implications for morphology-sensitive NLP in low-resource settings.