Artificial intelligence and machine learning
Research Article
Training Large Language Models to Handle Russian Verb–Noun Collocations
Vera Pavlovna Krasnobaeva
| ITMO University, St. Petersburg, Russia | |
|
krasnobaevavera2022@gmail.com |
Abstract. Non-compositional lexis— collocations, idioms and verb–noun constructions— remains a weak spot of large language models (LLMs), especially in morphologically rich languages such as Russian, where the boundary between compositional and non-compositional readings is graded. Russian verb–noun collocations were used as a controlled testbed for this boundary. New methodologically balanced dataset of 3,264 expressions is based on 479 verbs, with an equal number of collocations and compositional constructions and class balance for every verb, and define three subtasks over the same expressions: binary classification, single-verb paraphrasing, and gap filling. Seven multilingual and Russian-specialized LLMs are evaluated in zero-shot, one-shot and few-shot regimes, and two mid-size models are fine-tuned with LoRA in a multitask setting. Base models classify near chance and are unstable in generation; scale and language specialization help but do not close the gap. Fine-tuning raises classification accuracy from about 53% to about 80%, and on the generative subtasks a fine-tuned 8B model matches or surpasses much larger models. The paper extends a talk at the Dialogue-2026 conference; the dataset and code are publicly available. (Linked article texts in English and in Russian).
Keywords: verb–noun collocations, non-compositionality, idiomaticity detection, large language models, parameter-efficient fine-tuning, LoRA, paraphrasing, gap filling, Russian NLP
MSC-2020
68T50; 68T07, 91F20
For citation: Vera P. Krasnobaeva. Training Large Language Models to Handle Russian Verb–Noun Collocations . Program Systems: Theory and Applications, 2026, 17:5, pp. 3–42. (in Engl. In Russ.). https://psta.psiras.ru/2026/5_3-42.
Full text of bilingual article (PDF): https://psta.psiras.ru/read/psta2026_5_3-42.pdf (Clicking on the flag in the header switches the page language).
The article was submitted 11.10.2026; approved after reviewing 13.10.2026; accepted for publication 17.10.2026; published online 06.11.2026.
