Research seminar Artificial intelligence Machine learning Data analysis Computer science
- Date: September 30, 2026
- Speakers: Olga Mitrofanova, St. Petersburg University
Modern language models successfully cope with the generation of coherent texts, reproducing user-specified characteristics (language, style, structure, etc.).
At the same time, there are linguistic tasks that present difficulties for LLMs, in particular, modeling lexical-semantic relations in word combinations, differentiating synonymy and antonymy, modeling intraword relations, etc.
The report discusses solutions that improve the interpretability of LLM outputs. The materials of the report may be of interest both to linguists who use LLMs in their research and to applied mathematicians who develop new neural network architectures.
References:
Astapenka, L., & Mitrofanova, O. (2026). Prediction of Lexical-Semantic Relations In Belarusian: From Word2Vec to LLM. In Computación y Sistemas, 30(1)
Athugodage, M., & Mitrofanova, O. (2025, June). The Influence of Language Typological Features on Neural Summarization Performance. In International Conference on Internet and Modern Society (pp. 132-144). Cham: Springer Nature Switzerland.
Bakaev, M., Gorovaia, S., & Mitrofanova, O. (2025). Who will author the synthetic texts? Evoking multiple personas from large language models to represent users’ associative thesauri. In Big Data and Cognitive Computing, 9(2), 46.
Mitrofanova, O., Iurevtseva, P., & Bakaev, M. (2025, October). What Do LLMs Know About Human Emotions? The Russian Case Study. In International Conference on Speech and Computer (pp. 129-144). Cham: Springer Nature Switzerland.
Rodionova, D. K., & Mitrofanova, O. A. (2026). Representation of Intraword Syntagmatic Relations in Vector Language Models. In Automatic Documentation and Mathematical Linguistics, 60(Suppl 3), 353-359.