From words to visuals: a transformer-based multi-modal framework for emotion-driven tourism analytics

Calderón-Fajardo, Víctor; Rodríguez-Rodríguez, Ignacio; Puig-Cabrera, Miguel

http://hdl.handle.net/10400.1/28375

Utilize este identificador para referenciar este registo.

Nome:	Descrição:	Tamanho:	Formato:
s40558-025-00334-2.pdf		3.72 MB	Adobe PDF	Ver/Abrir

Contacte-nos

Autores

Calderón-Fajardo, Víctor

Rodríguez-Rodríguez, Ignacio

Puig-Cabrera, Miguel

Resumo(s)

Traditional tourism analytics have primarily relied on isolated sentiment analysis and image processing techniques, often failing to capture the subtle interaction between textual expressions and visual aesthetics inherent in tourist experiences. This study addresses these limitations by proposing a novel multi-modal framework that transforms textual reviews into AI-generated images using standardized prompts, thereby converting affective signals into explicit visual features. Leveraging stateof-the-art models—such as Distilled Bidirectional Encoder Representations from Transformers (DistilBERT) for fine-grained emotion recognition and Contrastive Language–Image Pre training (CLIP) for semantic extraction of visual attributes— our approach maps complex sentiments onto interpretable visual characteristics, integrating explainable features to uncover the underlying structure in tourist perceptions. This approach enhances classification performance and provides a transparent mechanism for understanding how distinct emotional states correspond to specific visual cues. Experimental evaluations on a dataset encompassing four diverse tourist destinations—Berlin, Dublin, Cairo, and Málaga—demonstrate high classification accuracy and robust correlations between text-derived emotions and image-based features, close to more powerful embedding methods. Significant correlations were observed between emotions and visual features, e.g., brightness and contentment, as well as between entropy and shame, indicating that our method efficiently captures the affective resonance between visual and textual modalities. Our findings underscore the transformative potential of converting textual sentiment into visual representations to facilitate more accurate, interpretable, and actionable analytics in the tourism sector. This framework suggests promising avenues for dynamic destination characterization, informed marketing strategies, and enhanced urban planning initiatives, laying the foundation for future advancements in multimodal tourism analytics.