Integration of Descriptive Sensory Analysis and Large Language Models for Construction of a Sensory Wheel for Sufu (Chinese Fermented Soybean Curd)

Sufu is known for its soft, creamy, “cheese-like” texture and strong flavor. It is commonly classified into types such as red, white (plain), gray, rose and flavored. In this study, 66 commercial sufus representing these types were evaluated using descriptive sensory analysis to gather descriptors for constructing a sufu sensory wheel. Preliminary lexicons were generated and refined by Frequency ( F ), Intensity ( I ) and Geometric Mean ( M ) values, resulting in a final set of 102 descriptors: 16 for appearance, 36 for odor/aroma, 20 for mouthfeel/texture, and 30 for taste/flavor. These attributes were subsequently grouped using two methods: DeepSeek semantic analysis identified 22 subcategories, while Agglomerative Hierarchical Clustering (AHC) identified 17 subcategories, leading to two distinct visualizations of the sufu sensory wheel. The results showed that DeepSeek, by analyzing the precise semantics of words and phrases, provides a more objective, language-driven classification that is less influenced by subjective evaluation metrics and human bias. In contrast, AHC proved more susceptible to inconsistencies in panel judgments, which could lead to less reliable or semantically less intuitive groupings. The sufu sensory wheel is particularly useful for panel training, product education, and communication with manufacturers and non-expert audiences. It may also serve as a reference for future sensory studies or product development in the commercial sufu industry. Looking ahead, a hybrid approach that integrates both computational methods with the expertise of sensory scientists could effectively combine empirical grounding with semantic precision, yielding more robust and comprehensive sensory wheel frameworks for sufu.