Anna Chinni¹, Kushin Mukherjee², Karen B. Schloss¹; ¹University of Wisconsin - Madison, ²Stanford University

A central goal in research on information visualization is to identify generalizable principles and metrics that predict people’s ability to interpret visualizations, transcending any one feature or chart type. Previous work within the domain of color semantics developed semantic distance as a metric to predict performance on assignment inference, the process by which people infer the mappings between perceptual features and concepts in visualizations. Semantic distance is computed from the probability of one assignment over the alternative(s). For color, semantic distance predicted accuracy in assignment inference, independent of perceptual distance and color-concept association strength (Schloss et al., 2021). In this study, we tested the generalizability of semantic distance as a predictor of interpretability by studying assignment inference for a different type of perceptual feature: visual texture. During the experiment, participants viewed a series of bar graphs, in which the textures of the bars represented different concepts (e.g., bubbly and stripey textures representing the concepts ocean and field). On each trial, participants were given a target concept (e.g., ocean) and they reported which textured bar mapped onto that concept. Across trials, the texture sets spanned a wide range of semantic distances and varied systematically with respect to the concepts they represented. We found that semantic distance correlated strongly with accuracy in assignment inference (e.g., r = .88, p < .001 for the concepts ocean and field). A logistic mixed effects model showed that semantic distance predicted accuracy independent of perceptual distance (estimated by the Euclidean distance between textures in an embedding space constructed from triadic similarity judgements (Chinni et al., VSS-2025)) or association strength for the optimal texture choice. These results support the possibility that semantic distance is a generalizable metric for predicting performance in assignment inference, which can be used broadly to design visualizations that facilitate communication.

Acknowledgements: This work is supported in part by the National Science Foundation Grant BCS-2419493