What the paper studied
This paper addresses a persistent gap in how hospitality operators use guest reviews: while reviews are a rich source of operational intelligence, most analytics approaches—manual reading or simple sentiment scoring—fail to pinpoint which specific aspects of the guest experience need improvement and why. The researchers propose a dual-model AI framework that combines BERT’s supervised classification with Gemini’s generative reasoning to extract detailed, aspect-level insights from review text. Instead of merely assigning positive or negative scores, the system identifies the precise elements driving dissatisfaction—such as service, accessibility, or cultural experience—and generates actionable recommendations. The framework was validated using TripAdvisor reviews of the Archaeological Site of Mystras, a UNESCO World Heritage Site, and includes a visualization dashboard designed for practical use by hospitality managers.
Key findings
- BERT efficiently classifies reviews into structured categories at scale, but its reliance on star ratings can cause it to overlook nuanced complaints embedded in otherwise positive reviews. For example, a four-star review with a significant accessibility complaint may still be classified as positive.
- The Gemini generative model layer addresses this limitation by analyzing the full context of each review, surfacing negative sentiment threads even when the overall rating is high. This dual-model approach captures qualified positives and criticisms that single-model systems often miss.
- The framework’s aspect-based analysis independently scores key dimensions such as accommodation quality, service responsiveness, cultural experience, pricing perception, and accessibility. The generative model then produces plain-language recommendations tailored to the specific aspects showing negative sentiment.
- The visualization dashboard enables managers to track patterns in sentiment, aspect categories, and severity over time, making the insights accessible to non-technical staff and suitable for regular operational review.
Why it matters for hospitality
Traditional review analytics often miss the actionable details that drive guest dissatisfaction, limiting the ability of operators to make targeted improvements. By moving to aspect-level sentiment analysis with AI-generated recommendations, hospitality properties can identify and address specific operational weaknesses more effectively. This approach helps bridge the gap between what leading properties can extract from guest feedback and what average properties achieve, creating a compounding feedback loop: better problem identification leads to faster fixes, improved scores, more reviews, and richer data for future analysis.
Practical takeaways
- Adopting a dual-model AI framework can reveal hidden or nuanced complaints within generally positive reviews, improving the detection of operational issues that might otherwise go unnoticed.
- Aspect-level sentiment analysis provides more granular and actionable insights than overall sentiment scores, enabling targeted interventions in areas like accessibility or service responsiveness.
- Generative AI models can translate complex sentiment data into clear, operational recommendations that managers can act on without needing data science expertise.
- Visualization dashboards make it feasible for non-technical staff to monitor sentiment trends and aspect-level issues regularly, supporting data-driven decision-making.