Prof. Dr. Ualsher Tukeyev,
National University of Uzbekistan named after Mirzo Ulugbek, Uzbekistan
LLM FOR ASPECT-BASED SENTIMENT ANALYSIS OF UZBEK SERVICE SECTOR REVIEWS
The rapid growth of digital platforms has transformed online consumer reviews into an important source of social-behavioural data, reflecting how citizens evaluate services, express satisfaction or dissatisfaction, and communicate expectations toward businesses and institutions. In emerging digital societies, such data can support evidence-based service improvement and provide valuable insight into patterns of public perception. However, for low-resource languages such as Uzbek, the practical use of online review data remains limited by the lack of large, high-quality annotated datasets and robust language technologies. This talk addresses that gap by presenting a large language model (LLM)-driven framework for Aspect-Based Sentiment Analysis (ABSA) of Uzbek online service reviews, with a focus on the banking, retail, and HoReCa sectors. The proposed direction is closely aligned with the conference’s interest in applying computational and AI-based methods to social-behavioural phenomena. Unlike conventional sentiment analysis, which reduces reviews to broad positive or negative labels, ABSA enables a more fine-grained understanding of how users judge specific aspects of service experience, such as staff behaviour, response speed, pricing, delivery quality, product reliability, and digital interaction quality. This makes ABSA especially relevant for modelling consumer perceptions as complex social signals rather than simple emotional reactions. In multilingual and low-resource contexts, such modelling is not only a technical challenge but also a social one, because it determines whether digitally expressed public opinion can be transformed into structured knowledge for decision-making and service redesign. The Uzbek case is particularly important because it represents a growing digital market where user-generated content is increasing, while computational resources for language understanding remain underdeveloped.
The proposed research framework is based on the collection and processing of more than 30,000 Uzbek-language consumer reviews from the Sharh platform. To overcome the bottleneck of manual annotation, the study employs open-source LLMs, including models such as Llama 3 and Mistral, as annotation assistants for aspect extraction and sentiment labelling.
