Customer Review Topic Modelling and Emotion Analysis
What drives negative reviews for a UK wellness centre chain? Topic modelling, emotion classification and an instruction-tuned LLM applied to 27,585 reviews from two public review platforms.
The Challenge
Topic modelling of 2,031 angry reviews traced the anger mainly to four issues: cleanliness, staff conduct, equipment and membership fees.
A national wellness centre chain receives thousands of reviews across public platforms. Star ratings show that customers are unhappy, but not why, or where. The task was to turn unstructured review text into specific, location-aware themes that operations teams can act on.
The work combined classical NLP with transformer models and an instruction-tuned LLM, and compared what each approach surfaced.
Approach
Results
Negative reviews numbered 2,422 on one platform and 3,396 on the other. Among angry reviews, the largest themes were cleanliness of toilets and changing rooms (334 reviews), staff rudeness (326), equipment (271) and membership fees and cancellation (260). The first topic-modelling run on negative reviews pointed the same way: showers and changing-room cleanliness first, then broken equipment.
Seven locations appeared in both platforms' top-20 lists for negative reviews, giving a short list of sites where problems are consistent rather than platform-specific.
Limitations: the first BERTopic run used the first 1,000 negative reviews rather than a random sample, and the LLM step ran on 50 angry reviews only. The emotion model labelled about 1,700 negative reviews as "joy", a likely misclassification. LLM-generated suggestions are generic and were not validated.
How it was tested
- Reviews filtered to English, de-duplicated and stripped of personal names and brand terms before analysis.
- Three approaches compared on the same reviews: BERTopic, Gensim LDA and an instruction-tuned LLM.
- The emotion model labelled about 1,700 negative reviews as "joy", a likely misclassification.
- The first BERTopic run used the first 1,000 negative reviews rather than a random sample.
- The LLM step ran on 50 angry reviews only, and its suggestions are generic and were not validated.