LLM in sentiment analysis of fuzzy Russian perfume consumer reviews

LLM in sentiment analysis of fuzzy Russian perfume consumer reviews

Dolmatova O.V. (PSU, Pyatigorsk, Russia)

Abstract

The paper investigates the potential of large language models (LLM) in sentiment analysis for less obvious perfume consumer reviews, i.e., those with no or little emotionally charged vocabulary. Not only the polarity of reviews but the decision-making process is of prominent interest. The dataset consists of one hundred consumer reviews from fragrantica.ru. The author chose large language models Alice AI (developed by Yandex) and GigaChat (developed by Sberbank) as research tools for their ability of providing detailed feedback explaining the bias of their solution in a natural, human-friendly form. Both models demonstrated sound analysis, high predictive power, and were largely unanimous in their decisions. Discrepancies in assessments and disagreements with the opinions of the review authors were minimal; these cases appear to be particularly valuable both for the principles of decoding figurative perfume reviews and for establishing the possible blind spots of models, and, accordingly, their possible improvement if necessary. The study revealed the following patterns of sentiment analysis fulfillment: if there are more than one concept the model selects one considered to be the main one, assigns a certain feature to it and pulls the interpretation of the remaining concepts to the initial one, creating a holistic, consistent concept; such main concepts may differ among different models; if the models fail to detect the concepts clearly expressed in words and phrases, the method by contradiction is used - the absence of a clear negation is perceived as positive or neutral; concepts available for analysis tend to receive a positive interpretation.

Keywords

large language models; Alice AI; GigaChat; sentiment analysis; consumer review.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 3, part 1, 2026, pp. 197-208

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

DOI: 10.15514/ISPRAS-2026-38(3)-12

For citation

Dolmatova O.V. LLM in sentiment analysis of fuzzy Russian perfume consumer reviews. Proceedings of the Institute for System Programming, vol. 38, issue 3, part 1, 2026, pp. 197-208 DOI: 10.15514/ISPRAS-2026-38(3)-12.

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