A local large language model was used to extract quantitative preference data from hotel reviews and construct a fuzzy cognitive map capable of predicting guest satisfaction scores.
Researchers fed unfiltered TripAdvisor reviews into Qwen2.5-32B, a 32-billion-parameter model that runs locally rather than through a cloud API. The model accepted hotel-related concepts as prompt inputs and returned quantitative weights representing how much reviewers cared about each factor. Those weights were then used to train a fuzzy cognitive map — a graph-based reasoning model that encodes causal relationships between concepts. Testing focused on Greek-language reviews, where the resulting map took a star topology that surfaced which factors most influenced overall satisfaction.
The external validation step is the part worth watching: the fuzzy cognitive map was asked to correlate its predicted satisfaction with the numeric star rating from the original review — data it never saw during inference. That it could do so meaningfully suggests LLM-extracted weights carry enough signal to anchor a downstream reasoning model, not just describe sentiment in broad strokes. For applied research teams, it also lowers the barrier to building domain-specific cognitive models without hand-labeled training data.
Fuzzy cognitive maps have been around since the 1980s, but feeding them with LLM-extracted weights rather than expert elicitation is a newer wrinkle — one that trades interpretability for scale. Whether Qwen2.5-32B's outputs are consistent enough across runs to make the resulting maps reliable is a question the paper flags but does not fully close.