A comparative analysis of verb ellipsis generation by large language models (LLMs) and humans

A comparative analysis of verb ellipsis generation by large language models (LLMs) and humans

Naidenova X.A. (MMA, St. Petersburg, Russia)
Bulykina E.S. (SPbPU, St. Petersburg, Russia)
Parkhomenko V.A. (SPbPU, St. Petersburg, Russia)
Martirova T.A. (MMA, St. Petersburg, Russia)

Abstract

This paper aims to conduct a comparative analysis of how large language models (LLMs) and humans generate verbal ellipsis. The relevance of this research stems from the need to understand how neural networks can reproduce complex and context-dependent linguistic phenomena that are characteristic of natural speech.In the course of the research, a serial experiment was conducted in which the GigaChat, YandexGPT and Gemini models transformed complete sentences into elliptical ones. The RAG (c) architecture was used to correct generation errors and improve contextualization. It is noted that LLMs, unlike humans, demonstrate stricter formal adherence to instructions on verb exclusion. The results show the similarity of LLM and human strategies. The research raises the question of the validity of existing metrics for evaluating similar linguistic tasks.

Keywords

ellipsis; large language models; RAG.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 3, part 4, 2026, pp. 145-156

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

DOI: 10.15514/ISPRAS-2026-38(3)-52

For citation

Naidenova X.A., Bulykina E.S., Parkhomenko V.A., Martirova T.A. A comparative analysis of verb ellipsis generation by large language models (LLMs) and humans. Proceedings of the Institute for System Programming, vol. 38, issue 3, part 4, 2026, pp. 145-156 DOI: 10.15514/ISPRAS-2026-38(3)-52.

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