A Framework for Semantic Annotation of Russian-Language Tables Using Deep Learning

Structured tabular data remains a cornerstone of information representation, serving critical roles in knowledge sharing, analytical workflows, and cross-domain interoperability. Despite their ubiquity, automated interpretation of tables is hindered by their inherent structural diversity, including non-standard formats, context-dependent semantics, and language-specific nuances. While recent advancements in pretrained language models have contributed to advances in table understanding, existing research is mainly focused on English-oriented datasets, resulting in under-explored low-resource languages such as Russian. To bridge this gap, we introduce RuTaBERT, a deep learning framework for semantic column annotation of Russian tables using transformer-based multilingual models. Our framework maps tabular data to semantic type labels from DBpedia, addressing linguistic features through domain-adaptive fine-tuning. The training utilizes a curated dataset of 3.2 million columns derived from the Russian Wikipedia corpus, preprocessed to reduce noise and lexical ambiguity. For this purpose, we develop a new LoReTA framework for automating the preparation of tabular data, including their automatic labeling based on table headers, which is suitable for various languages. RuTaBERT integrates context-enriched table linearization and a contrastive pre-training strategy to handle linguistic specificity and data sparsity. Experimental evaluations demonstrate RuTaBERT’s advantage, achieving a high Micro F1 score of 97.4% on an established benchmark, far surpassing baselines while demonstrating robustness on rare semantic types and computational efficiency. By transcending language barriers, RuTaBERT unlocks opportunities for knowledge graph enrichment, data integration, and automated analytics in low-resource language contexts.