Large Language Model-based Topic-Level Sentiment Analysis for E-Grocery Consumer Reviews

Topic-level sentiment analysis is becoming increasingly important in understanding customer opinions. This study comprehensively evaluates BERT-based clustering methods for the topic detection task of the LLM-based topic-level sentiment analysis framework. The focus is comparing Fuzzy C-Means for soft clustering with K-Means, DBScan, and HDBScan for hard clustering. Experiments on three benchmark datasets show that each method performs differently depending on the dataset’s structure and semantic complexity. While Fuzzy C-Means excels in generating semantically coherent topics in structured data, HDBScan captures diverse topic spaces in heterogeneous content, and K-Means offers the best balance across metrics. The framework is applied to real-world consumer reviews from an Indonesian e-grocery platform and uncovers that 31.7% of all negative sentiment concentrates around the “shopping experience” topic despite an overall positive sentiment trend. This study offers methodological insight and practical guidance for selecting clustering techniques in LLM-based topic-level sentiment analysis and demonstrates their business value in digital consumer analytics.