Abstract. The Kazakh language is designed so that grammatical information is distributed along a chain of affixes, rather than concentrated in individual words. This in itself creates difficulties for automated systems. The situation is further complicated by the constant influx of Russian and English borrowings, which are partially dislodged from the phonological logic of the language, and as a result, the morphological analyzer is faced with not one type of difficult forms, but with several fundamentally different ones. Existing systems evaluate the performance of NLP models in an aggregated manner, without distinguishing the types of anomalies by the mechanism of their impact on the analysis, which does not allow targeted management of processing. The subject of this study is the relationship between the types of morphological errors and performance indicators of transformative NLP models for the Kazakh language. The aim of the work is to develop a formalized taxonomy of morphological anomalies and integrate it into a hybrid NLP architecture as a routing mechanism for analytical components. The study uses formal linguistic analysis, the corpus method based on the Universal Dependencies Kazakh-KTB (1,047 sentences), computational modeling and ablative analysis. The KazMorphCorpus-2026 hybrid architecture combines rule-based FST analysis, CRF disambiguation, the KazRoBERTa transformer module and the MFRN feature morphological compatibility verification module. Based on the results of the study, a five–class taxonomy of morphological anomalies was proposed – borrowings (BOR), affixal complexity (AFC), segmentation disorders (SEG), conflicts of grammatical features (AGR) and neutral class (NONE), integrated into the system as a control routing mechanism. In the test sample, the system reaches Accuracy = 87.4% and Macro-F1 = 0.86; the largest increase in quality was recorded for the AGR (ΔF1 = +0.14) and AFC (ΔF1 = +0.12) classes. The experiment confirmed that different types of morphological anomalies have different effects on the operation of the transformer model, and this difference is of practical importance. Systems that diagnose the type of anomaly prior to analysis and direct the token to a suitable component produce a result that is easier to interpret and easier to improve purposefully.

Keywords: morphological error taxonomy, morphological routing, hybrid NLP architecture, transformer models, KazRoBERTa, FST, CRF, MFRN, agglutinative language.