LLM-Assisted Weak Supervision for Low-Resource Kazakh Sequence Labeling: Synthetic Annotation and CRF-Refined NER/POS Models

Abstract

Kazakh sequence labeling is constrained by limited annotated resources, while its agglutinative morphology and productive suffixation increase data sparsity and exacerbate label inconsistency in part-of-speech (POS) tagging and named entity recognition (NER). This paper proposes an LLM-assisted weak supervision framework in which a large language model generates synthetic token-level annotations that are subsequently filtered using confidence-based criteria and combined with a smaller manually verified subset to train Transformer-based sequence taggers with Conditional Random Field (CRF) decoding. The pipeline unifies corpus construction, weak-label generation, quality filtering, word-to-subword alignment, and CRF-refined structured prediction into a reproducible workflow. Experimental results show that contextual encoders and structured decoding provide strong performance for Kazakh POS and NER, while the proposed training design enables efficient convergence with diminishing returns beyond moderate epoch budgets. Error-slice analysis indicates that residual errors are concentrated in rare tokens, morphologically complex long words, longer sentences, and the ORG entity class. Overall, the findings support the use of LLM-assisted weak supervision as a scalable strategy for low-resource Kazakh sequence labeling when synthetic labels are controlled through filtering and refined by structured decoding.