Study of Semantic Representation of Tabular Data for Fact Verification

Study of Semantic Representation of Tabular Data for Fact Verification

Poddubny I.A. (IDSTU SB RAS, Irkutsk, Russia)
Dorodnykh N.O. (IDSTU SB RAS, Irkutsk, Russia)
Armaev A.S. (IDSTU SB RAS, Irkutsk, Russia)

Abstract

Tables are widely used in various fields, from analytics to management, but their automatic interpretation is hampered by structural heterogeneity and the lack of explicit semantics. This paper proposes an approach to table reasoning based on a compact semantic representation of tables and the generation of executable Pandas queries for them. The representation is a table data schema in XML format, enriched with data types, semantic column roles, and a contextual description, which preserves structural dependencies while significantly reducing the data size of the original table. Experiments on the TabFact dataset showed that the model with the LoRA adapter achieves the best balance between accuracy (80.15%) and code executability (95.38%). When comparing the model with the baseline solution without special error correction, the model with the LoRA adapter achieves better accuracy. The paper also provides a detailed analysis of the errors that prevent the model from demonstrating final results similar to the baseline solution with error correction. To confirm the transferability of the proposed approach to other data, additional experiments were conducted on the WikiTableQuestions dataset in the context of the task of answering questions on tabular data.

Keywords

tabular data; table reasoning; table fact verification; question-answering system; tabular representation; semantic table representation; query execution.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 97-122

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(6)-7

For citation

Poddubny I.A., Dorodnykh N.O., Armaev A.S. Study of Semantic Representation of Tabular Data for Fact Verification. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 97-122 DOI: 10.15514/ISPRAS-2026-38(6)-7.

Full text of the paper in pdf (in Russian) Back to the contents of the volume