Qwen2.5-32B turns TripAdvisor reviews into fuzzy cognitive maps for satisfaction modeling
Local LLM extracts quantitative data from hotel reviews to predict star ratings with a causal map
Alexis Kafantaris published a new paper showing how a local large language model—specifically Qwen2.5-32B—can be used to develop fuzzy cognitive maps (FCMs) from raw textual data. The approach bypasses traditional preprocessing steps, instead feeding unfiltered TripAdvisor hotel reviews directly into the LLM. The model then extracts quantitative causal relationships between entities (like service quality, location, or cleanliness) and outputs data suitable for constructing an FCM—a graphical model that captures how factors influence each other.
When applied to a set of Greek hotel reviews, the method produced a star-topology FCM that clearly indicated the preferences of reviewers. Kafantaris validated the model by testing whether the FCM could predict an external metric—the star rating of a review—that was never used during training. The results showed a strong correlation between the FCM's predicted satisfaction and the actual star rating, demonstrating that local LLMs can bridge the gap between unstructured text and structured causal models, with potential applications in sentiment analysis, customer feedback, and decision support systems.
- Qwen2.5-32B extracted quantitative causal data directly from raw TripAdvisor reviews without preprocessing.
- The resulting fuzzy cognitive map (FCM) used a star topology to represent reviewer preferences from Greek hotel feedback.
- External validation showed the FCM could predict star ratings—an outcome outside the LLM's inference scope—with high correlation.
Why It Matters
Shows how local LLMs can turn unstructured text into actionable causal models, enabling data-driven decisions from reviews.