BENCHMARKING DEEP LEARNING MODELS FOR FEW-SHOT CLASSIFICATION OF KAZAKH NEWS TEXTS UNDER LIMITED ANNOTATION

Abstract

Deep learning methods and transformer architectures have fundamentally reshaped the field of natural language processing; however, these advancements remain unevenly distributed. While high-resource languages like English benefit from large-scale benchmarks, robust text classification for low-resource languages, such as Kazakh, continues to pose a significant challenge. This paper presents a comparative analysis of the classical Word2Vec embedding, BiLSTM recurrent models, BERT and XLM-R transformers, the generative mT5 model, alongside classical and statistical baselines for few-shot Kazakh news text classification under limited annotation constraints. Experiments were conducted on the KazNews dataset, curated for few-shot classification from tengrinews.kz and inform.kz articles, as well as parallel HuffPost corpora. The datasets comprise five categories (sports, politics, business, travel, etc.). The experimental findings in 1-shot and 5-shot settings demonstrate that no single model consistently dominates across all datasets and regimes: Word2Vec achieved the best performance on the HuffPost corpora in the 5-shot regime (0.48 and 0.53); mT5-Seq2Seq outperformed others on KazNews under 5-shot (0.71); while BERT remained competitive on the long-text KazNews dataset in the 5-shot setup. The authors also observed a sharp performance gain for the mT5-Seq2Seq model in the 5-shot setup. Conversely, none of the evaluated models yielded acceptable quality metrics in the 1-shot regime.