Measuring Individual Differences in Meaning: The Supervised Semantic Differential

The Supervised Semantic Differential (SSD) is a method for measuring individual differences in how people construe the same concept in open-ended language based on relatively small text samples. SSD represents each participant’s concept-related language use as a personal concept vector, relates these representations to an external quantitative variable, and recovers an interpretable semantic gradient describing how meaning shifts across the scale of that variable. Because the recovered gradient is expressed in the original embedding space, it can be interpreted through semantically related words, clusters, and representative text excerpts. We evaluated SSD across seven corpora of short essays written by 1,736 Polish adults; each paired with a corresponding questionnaire measure. Across corpora, SSD recovered statistically reliable semantic gradients with adjusted R^2 values ranging from .03 to .14, with clear qualitative interpretations that varied in coherence and polarization depending on the variance explained. To assess construct validity, we additionally applied SSD to a lexical-norm dataset containing ratings for 4,905 Polish words on eight affective and psycholinguistic dimensions. In this setting, SSD recovered established dimensions such as valence, dominance, concreteness, and age of acquisition with strong quantitative fit and highly interpretable semantic poles. To assess nomological validity, we compared the association patterns of questionnaire-based and SSD-based scores with demographic and behavioral variables; SSD generally preserved the broader correlational structure of the original constructs, although in attenuated form. Finally, we provided a statistical power analysis to assess what amount of text records is needed to achieve proper power. Taken together, these findings suggest that SSD provides a practical and interpretable framework for studying individual differences in meaning from open-ended text. More broadly, the method offers a way of linking free-response language to psychologically meaningful semantic structure at sample sizes typical of psychological research.