Dependence of Text Intrinsic Dimension on Language

Dependence of Text Intrinsic Dimension on Language

Kuchina M.A. (MIPT, Dolgoprudny, Moscow Region, Russia; ICS RAS, Moscow, Russia)
Kildyakov A.S. (ICS RAS, Moscow, Russia)
Papay I.D. (MIPT, Dolgoprudny, Moscow Region, Russia)

Abstract

This work investigates the dependence of the intrinsic dimension of a text on its language. The hypothesis that the intrinsic dimension is invariant under translation is tested. The study is conducted on a multilingual corpus consisting of texts with the same content, translated into several languages. For each text, the intrinsic dimension is computed based on token embeddings obtained from a language model, followed by a comparative analysis of the resulting values. The analysis rejects the hypothesis of invariance under translation, confirming that the intrinsic dimension depends on the language. It is shown that synthetic languages with rich morphology give higher values than analytic ones.

Keywords

intrinsic dimension; translation invariance; text embeddings; natural language processing.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 243-258

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

DOI: 10.15514/ISPRAS-2026-38(6)-16

For citation

Kuchina M.A., Kildyakov A.S., Papay I.D. Dependence of Text Intrinsic Dimension on Language. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 243-258 DOI: 10.15514/ISPRAS-2026-38(6)-16.

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