D.S. Larionov1, E.N. Nikitina2, I.V. Smirnov3

1-3 Federal Research Center "Computer Science and Control" of the RAS (Moscow, Russia)

3 RUDN University (Moscow, Russia)

1 dslarionov@isa.ru, 2 yelenon@mail.ru, 3 ivs@isa.ru

When extracting arguments from emotive predicates, there is a need for automatic semantic (thematic) classification of the extracted arguments. The aim of this study is to investigate the feasibility of using large-scale language models (LLM) for automatic argument classification, and analyze the relationship between certainty and the quality of LLM results. Using the topic "Medicine and Healthcare," it has been experimentally demonstrated that the Claude Sonnet 4.5 language model enables high-quality argument classification. Furthermore, integrating explicit certainty assessment into prompts yields a statistically significant signal for identifying potentially unreliable predictions. The results of this study can be applied to systems for analyzing the emotional attitudes of social media users toward given topics and events.

Larionov D.S., Nikitina E.N., Smirnov I.V. Exploring large language models for semantic classification of arguments of emotion predicates. Highly Available Systems. 2026. V. 22. No 1. P. 12?16. DOI: https://doi.org/10.18127/j20729472-202601-02 (in Russian)

1. Lyashevskaya O., Kashkin E. FrameBank: A Database of Russian Lexical Constructions. AIST 2015. Springer, 2015. P. 350-360.
2. Anthropic. The Claude 3 Model Family: Opus, Sonnet, Haiku. Technical Report. 2024.
3. Xiong M. et al. Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs. arXiv preprint arXiv:2306.13063. 2023. DOI: 10.48550/arXiv.2306.13063
4. Wei J. et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems. 2022. V. 35. P. 24824-24837.
5. Smirnov I.V., Larionov D.S., Nikitina E.N., Kazachonok G.A. LLM for Semantic Role Labeling of Emotion Predicates in Russian. Supercomputing Frontiers and Innovations. 2025. V. 12. No 3. P. 31-46. DOI: 10.14529/jsfi250303