Character N-Gram Features for Kazakh POS Tagging Under Strict OOV Evaluation

Abstract

We present a compact ablation study for Kazakh part-of-speech (POS) tagging under a strict out-of-vocabulary (OOV) protocol. From the NLA-NU Kazakh Dependency Tree-bank (kdt-NLANU, 934,640 tokens), we build a deduplicated benchmark of 48,277 unique word-Universal part-of-speech (UPOS) pairs and split by word form, ensuring zero train-test form overlap. A central processing unit (CPU)-only logistic regression (LR) model using character n-grams and lightweight contextual features reaches 87.13% accuracy and a 0.569 macro-averaged F1 score (macro-F1) on a 13-class OOV test set. The ablation shows that the saga solver, character unigrams, and class filtering are more reliable than standard balanced class weighting, which fails when classes have fewer than about ten instances. We also report a transformer reference point, but mark it as non-comparable because it uses a sentence-level split rather than the OOV protocol.

More by Iskander Akhmetov

• A Lightweight Extractive Compression Pipeline for Long-Context Retrieval-Augmented GenerationComputación y Sistemas · 2026
• KazakhStopEval: Stop-Word Inventories, Validation Benchmarks, and NLP Evaluation Artifacts for Kazakh Library OCR CollectionsMendeley Data · 2026
• Hierarchical Multi-Agent Extractive Framework for Long-Document Mind Map GenerationComputación y Sistemas · 2026
• Accuracy-Efficiency Trade-offs in Automated Cervical Cytology Classification: A Comparative Evaluation of ViT, DeiT, PiT, and Swin Transformer on SIPaKMeDArtificial Intelligence in Applied Sciences · 2026 · Citations: 1