Combining Classical and Transformer-based Approaches for Text Classification and Topic Modeling of the Tatar Language

Abstract

This article presents an experimental study on text classification methods for Tatar, a low-resource Turkic language. Using a news corpus of 153,162 documents, we systematically compare classical approaches (TF-IDF+LR, FastText+LR) against transformer-based architectures (mBERT, XLM-R, DistilBERT) across three tasks: binary (news vs. analytics), multi-class (9 categories), and multi-label (13 categories). Our findings show that optimal model choice depends on task complexity. For binary classification, DistilBERT achieves the highest macro F1-score (0.9542), marginally outperforming other transformers, while TF-IDF+LR yields a comparable result (0.93). In multi-class tasks, transformers significantly surpass classical methods-mBERT attains a macro F1-score of 0.87, outperforming TF-IDF+LR by 16.9 percentage points. For multilabel classification, mBERT again achieves the best results (micro $\mathrm{F} 1=0.76$, subset accuracy $=0.34$), more than doubling classical method performance. To analyze thematic structure, we performed BERTopic modeling on a 10,000-document subset, identifying 46 interpretable topics. While 82.6% align with expert-assigned categories, 17.4% constitute emergent thematic clustersincluding regional news, transportation, agriculture, healthcare, and discursive elements-providing a more nuanced understanding of Tatar-language content. These results offer practical guidance for developing automated classification systems for Tatar media and contribute to advancing NLP methods for Turkic languages.

More by Mullosharaf Arabov

• Parameter-Efficient Adaptation of Large Language Models to a Low-Resource Agglutinative Language: A Corpus, Datasets, and a Reproducible Benchmark for the Bashkir LanguagearXiv · 2026
• A comprehensive study of Tatar text classification methods: from corpus creation and Tatar2Vec embeddings to benchmark establishmentModeling and Analysis of Information Systems · 2026 · Citations: 1
• A Comprehensive Analysis of Arabic Natural Language Processing Research: Trends, Topic Evolution, and Research Gaps -- A Bibliometric and Topic-Based StudyarXiv · 2026
• Large Language Models for Low-Resource Languages: A Conceptual Framework for an Electronic Explanatory Dictionary of the Tajik LanguagearXiv · 2026