Comparing Large Language Models for Aspect-Based Sentiment Analysis of Student Experience from Russian-Language Online Course Reviews

P. 320–337.

Kalicheva V., Kirina M.

The rapid expansion of the online education market has produced large volumes of unstructured data—course reviews containing detailed evaluations of specific educational components. Aspect-based sentiment analysis (ABSA) makes it possible to move from a holistic assessment of the educational experience to an analysis of its individual constituents, identifying particular aspects of learning and students’ attitudes toward each of them. This paper investigates the applicability of large language models to the task of quadruple extraction (aspect term—opinion—sentiment—aspect category) on Russian-language online course reviews within a single pipeline and without fine-tuning on annotated data. Using a corpus of 300 reviews across eight subject areas and a manually annotated sample of 100 reviews, three large language models (GPT-4.1 mini, YandexGPT 5.1 Pro, and DeepSeek V3.2) are compared under strict, partial, and semantic matching metrics. GPT-4.1 mini achieved the best results (quadruple F1 of 0.4716 under strict and 0.6126 under partial matching). It is shown that the exact-boundary matching procedure standard in aspect-based sentiment analysis is overly strict for large language models: discrepancies with the gold standard are largely formal rather than substantive, as shown mainly by the gain under partial matching. We also propose a typology of model errors arising from the morphosyntactic specificity of Russian.

Keywords: Russian language, Aspect-Based Sentiment Analysis, EdTech, Large Language Models (LLM), prompt engineering

In book

Speech and Computer. 28th International Conference, SPECOM 2026, Ohrid, North Macedonia, September 17–18, 2026, Proceedings, Part I

Springer, 2027.