Supervised Semantic Differential (SSD) Method

• SSD is a mixed quantitative–interpretive method that models connotative meaning as a bipolar semantic gradient learned from text–outcome associations in embedding spaces.
• The method applies PCA and linear regression techniques (OLS or ridge) to transform high-dimensional embeddings into interpretable gradients aligning with constructs like Evaluation, Arousal, and Dominance.
• SSD extensions include interaction modeling and cross-lingual analysis, providing diagnostic tools to assess gradient stability, semantic coherence, and corpus-specific biases.

Supervised Semantic Differential (SSD) is a mixed quantitative–interpretive method that models how semantic meaning varies with an external variable by estimating a semantic gradient—an oriented direction in an embedding space—and interpreting its poles through neighborhood retrieval, clustering, and representative texts. In contemporary formulations, SSD operationalizes the logic of Osgood’s Semantic Differential in embedding geometry: rather than prespecifying bipolar adjective scales, it learns a relevant bipolar axis directly from labeled text or lexical data, while retaining statistical testability and explicit semantic interpretation (Ostrowicki et al., 26 May 2026).

1. Conceptual foundations and relation to semantic differential

SSD inherits the core ambition of classical semantic differential methods: to represent connotative meaning along bipolar dimensions. In the classical formulation, semantic differential locates stimuli on dimensions such as evaluation, potency, and activity, or on adjective pairs such as good–bad and strong–weak. SSD preserves the idea of bipolar semantic structure, but replaces direct scale construction with supervised estimation from text–outcome associations in a vector space (Ostrowicki et al., 26 May 2026).

This shift places SSD between several neighboring traditions. It differs from unsupervised semantic-axis methods, such as axes defined by seed-word differences or antonym pairs, because the direction is learned from an observed variable rather than fixed in advance. It also differs from standard supervised text regression because the output is not merely a predictive model: the regression weights are transformed into a single interpretable semantic gradient whose poles can be audited semantically. In the cross-lingual literature, this has been explicitly tied back to Osgood’s classic triad by treating Evaluation/Valence, Potency/Dominance, and Activity/Arousal as supervised gradients recoverable from lexical norms (Sikora et al., 27 May 2026).

Within this family, SSD has been described both as a method for modeling how the connotative meaning of texts varies with a continuous individual-difference variable and as a framework for representing psychological constructs as directions in a shared semantic space. These formulations are compatible. In one usage, the goal is to recover a psychologically meaningful continuum within a corpus; in another, the goal is to place multiple constructs into a common coordinate system such as Valence, Arousal, and Dominance (VAD), enabling construct-level comparison across instruments, datasets, and research traditions (Plisiecki et al., 13 Mar 2026).

2. Core mathematical formulation

The canonical SSD setup assumes items paired with outcomes $y$. Each item $i$ is mapped to an embedding $x_i \in \mathbb{R}^D$ using a fixed embedding model. Published implementations have used SIF-weighted averaging of GloVe word vectors, L2 normalization, and removal of the top principal component to reduce anisotropy; in some settings the semantic unit is a whole document, whereas in the canonical PCV variant local contexts around a focal lexicon are aggregated into Personal Concept Vectors (Ostrowicki et al., 26 May 2026).

Because the embedding space is high-dimensional, SSD typically applies PCA before regression. One formulation defines centered document embeddings $X \in \mathbb{R}^{n \times d}$, takes the SVD $X = U \Sigma V^\top$, retains $K$ components, and computes unwhitened PCA scores

$S = X V_K \in \mathbb{R}^{n \times K}.$

A linear model is then fitted,

$y = \alpha \mathbf{1} + S \beta + \epsilon,$

with $\beta$ estimated by OLS; a ridge variant

$\hat{\beta} = (S^\top S + \lambda I)^{-1} S^\top y$

is also available, although the reported analyses in that formulation use OLS (Plisiecki et al., 13 Mar 2026).

A closely related formulation standardizes features before PCA, regresses the outcome on the reduced representation $i$0,

$i$1

and back-projects the coefficients to the original embedding space as

$i$2

where $i$3 is the PCA loading matrix and $i$4 contains the pre-PCA feature standard deviations. In that convention, the SSD score for an item is

$i$5

and a normalized version

$i$6

may be used to make scores comparable across datasets (Ostrowicki et al., 26 May 2026).

The underlying interpretation is the same across conventions. SSD estimates a direction in the original embedding space such that larger projections are associated with larger outcome values. Texts with large positive projections define one semantic pole; texts with large negative projections define the opposite pole. The method is therefore simultaneously geometric, statistical, and interpretive.

3. Dimensionality selection, interpretability, and diagnostic machinery

A central methodological issue in SSD is the choice of retained PCA dimensionality $i$7. One recent formulation treats this not as a purely variance-preservation problem but as a joint criterion over representation capacity, gradient interpretability, and stability across nearby values of $i$8. The cumulative variance explained is

$i$9

but predictive fit is not used as the selection target. Instead, the method emphasizes whether the learned gradient yields semantically coherent pole neighborhoods and whether the direction stabilizes as $x_i \in \mathbb{R}^D$0 changes (Plisiecki et al., 13 Mar 2026).

Interpretability is operationalized through retrieval and clustering at the poles. In the AI discourse case study, the top-100 neighbors per pole were retrieved by cosine similarity to the gradient, k-means clustering was applied with cosine distance, and the number of clusters was chosen by silhouette from the range $x_i \in \mathbb{R}^D$1. Cluster coherence and centroid alignment to the gradient were then aggregated into an interpretability score. Stability was defined through cosine similarity between normalized gradients at adjacent $x_i \in \mathbb{R}^D$2 values,

$x_i \in \mathbb{R}^D$3

with unit change

$x_i \in \mathbb{R}^D$4

After AUCK local neighborhood averaging and median smoothing, dimensionality was chosen by

$x_i \in \mathbb{R}^D$5

selecting the smallest $x_i \in \mathbb{R}^D$6 that attained the maximal joint score (Plisiecki et al., 13 Mar 2026).

Interpretability in SSD is not limited to cluster labels. Standard tools include nearest-neighbor retrieval around each pole, clustering of neighborhood words or texts, extraction of representative snippets, and visualization of document projections onto the gradient. In interaction SSD, the same toolkit is applied not only to the shared main gradient but also to interaction and conditional gradients. The resulting analysis treats semantic neighborhoods as model-dependent maps whose local organization is part of the substantive evidence, while formal block tests provide the confirmatory inferential layer (Ostrowicki et al., 26 May 2026).

This diagnostic emphasis also clarifies a recurrent methodological point: maximizing retained variance or regression fit alone can degrade semantic coherence. In the AI study, a high-$x_i \in \mathbb{R}^D$7 counterfactual that retained approximately $x_i \in \mathbb{R}^D$8 of variance produced diffuse, weakly structured clusters despite higher adjusted $x_i \in \mathbb{R}^D$9, which was taken as evidence that SSD should privilege stable and interpretable gradients over over-parameterized solutions (Plisiecki et al., 13 Mar 2026).

4. Major extensions of SSD

One major extension is interaction SSD, designed for moderation questions in which the meaning–outcome relationship itself may vary with a group, trait, or condition. With standardized outcome $X \in \mathbb{R}^{n \times d}$0, PCA scores $X \in \mathbb{R}^{n \times d}$1, and standardized moderator $X \in \mathbb{R}^{n \times d}$2, the model is

$X \in \mathbb{R}^{n \times d}$3

Here, $X \in \mathbb{R}^{n \times d}$4 captures the shared semantic association, $X \in \mathbb{R}^{n \times d}$5 the non-semantic main effect of the moderator, and $X \in \mathbb{R}^{n \times d}$6 the semantic interaction. Back-projection yields a main gradient and an interaction gradient,

$X \in \mathbb{R}^{n \times d}$7

and a conditional gradient at moderator value $X \in \mathbb{R}^{n \times d}$8,

$X \in \mathbb{R}^{n \times d}$9

Wald $X = U \Sigma V^\top$0 tests are used for the semantic and interaction blocks, and partial $X = U \Sigma V^\top$1 quantifies the residual variance explained by the interaction block after accounting for the semantic block and moderator main effect (Ostrowicki et al., 26 May 2026).

A second extension represents constructs in a shared semantic space. In that framework, SSD estimates construct-specific gradients and projects them onto theoretically motivated reference axes, particularly VAD. Given unit-length construct gradient $X = U \Sigma V^\top$2 and unit reference axis $X = U \Sigma V^\top$3, the coordinate on axis $X = U \Sigma V^\top$4 is the cosine similarity

$X = U \Sigma V^\top$5

This makes otherwise incommensurable psychological measurements semantically commensurate. In the GoEmotions application, an additional orthogonalization step removed a shared “generic emotionality” direction before projection into VAD space (Plisiecki, 26 May 2026).

A third extension is cross-lingual SSD. Here, separate supervised gradients are estimated in a shared multilingual embedding space obtained by orthogonal alignment of monolingual spaces. The cross-language comparison uses cosine similarity

$X = U \Sigma V^\top$6

as the primary alignment statistic. The literature introduces two permutation procedures—an alignment test for $X = U \Sigma V^\top$7 and a difference test for $X = U \Sigma V^\top$8—together with bootstrap intervals on Fisher’s $X = U \Sigma V^\top$9 scale. When gradients differ significantly, interpretation proceeds via a difference gradient

$K$0

whose poles are clustered to summarize residual semantic divergence (Sikora et al., 27 May 2026).

These extensions show that SSD is not restricted to a single regression backend or a single inferential regime. OLS and ridge appear in PCA-based formulations, whereas the cross-lingual work instantiates SSD with a PLS backend, including a single-component closed form used inside permutation procedures and multi-component PLS for final fits. A plausible implication is that SSD is best understood as a broader supervised-gradient paradigm whose common denominator is back-projection into an interpretable semantic direction, rather than a single fixed estimator (Sikora et al., 27 May 2026).

5. Empirical uses and substantive findings

SSD has been applied to short-form discourse, lexical affect norms, questionnaire text, and hate-speech annotation. In a case study on AI discourse, a corpus of $K$1 short posts written by Prolific participants was embedded with 300-dimensional Dolma GloVe vectors using SIF weighting with $K$2 and top-PC removal. The PCA sweep selected $K$3 for Admiration, yielding adjusted $K$4, $K$5, $K$6, and $K$7. The resulting gradient contrasted optimistic, collaborative framings of AI—innovation, partnership, empowerment—with distrustful and derisive discourse. For Rivalry, the sweep selected $K$8, but adjusted $K$9 and $S = X V_K \in \mathbb{R}^{n \times K}.$0, and the analysis concluded that no robust semantic alignment emerged (Plisiecki et al., 13 Mar 2026).

In work on shared semantic space, SSD was used to recover VAD reference axes from 13,915 English words with affective norms and then to project 27 GoEmotions categories and Big Five personality domains into that space. The VAD directions were strongly recoverable: Valence at $S = X V_K \in \mathbb{R}^{n \times K}.$1 had $S = X V_K \in \mathbb{R}^{n \times K}.$2 and adjusted $S = X V_K \in \mathbb{R}^{n \times K}.$3; Arousal at $S = X V_K \in \mathbb{R}^{n \times K}.$4 had $S = X V_K \in \mathbb{R}^{n \times K}.$5 and adjusted $S = X V_K \in \mathbb{R}^{n \times K}.$6; Dominance at $S = X V_K \in \mathbb{R}^{n \times K}.$7 had $S = X V_K \in \mathbb{R}^{n \times K}.$8 and adjusted $S = X V_K \in \mathbb{R}^{n \times K}.$9, all with $y = \alpha \mathbf{1} + S \beta + \epsilon,$0. GoEmotions categories organized in the expected way, with love, joy, and admiration at the positive end of valence and anger and fear at the negative end; Big Five domain placements were described as broadly coherent, whereas facet-level placements were more exploratory because they relied on sparse questionnaire text (Plisiecki, 26 May 2026).

Interaction SSD was demonstrated on the UC Berkeley Measuring Hate Speech corpus. The analysis used $y = \alpha \mathbf{1} + S \beta + \epsilon,$1 annotation records whose target was coded as people of color, with annotator racial identity coded as white versus POC and then standardized. Using 300-dimensional Common Crawl GloVe with SIF weighting, all-but-the-top correction, and a PCA sweep that selected $y = \alpha \mathbf{1} + S \beta + \epsilon,$2, the interaction model achieved adjusted $y = \alpha \mathbf{1} + S \beta + \epsilon,$3. The semantic block was strongly significant; white annotators had a small positive main effect on scores, $y = \alpha \mathbf{1} + S \beta + \epsilon,$4, $y = \alpha \mathbf{1} + S \beta + \epsilon,$5, $y = \alpha \mathbf{1} + S \beta + \epsilon,$6, $y = \alpha \mathbf{1} + S \beta + \epsilon,$7; and the interaction block was significant, $y = \alpha \mathbf{1} + S \beta + \epsilon,$8, $y = \alpha \mathbf{1} + S \beta + \epsilon,$9, with partial $\beta$0. The shared gradient contrasted slur-based contempt, violent eliminationist language, and vermin metaphors with diasporic civic discourse, reflexive engagement with race, and celebratory recognition of POC achievement. The interaction gradient captured smaller group-linked differences in which semantic cues predicted hate-speech ratings, and a robustness reanalysis with CC840B embeddings replicated the block significance with partial $\beta$1 (Ostrowicki et al., 26 May 2026).

Cross-lingual SSD has been used to compare affective meaning across Polish, English, and French. Using aligned multilingual embeddings, all fitted gradients for Valence, Arousal, and Dominance where available were statistically significant within each language at $\beta$2. Alignment tests with $\beta$3 permutations rejected $\beta$4 for every language pair and dimension. The bootstrap $\beta$5 intervals for EN–PL were Valence $\beta$6, Arousal $\beta$7, and Dominance $\beta$8; EN–FR and PL–FR showed weaker but still substantial alignment, especially for Valence. Qualitative interpretation of difference gradients suggested that Valence was mostly shared, whereas Arousal and Dominance contained more interpretable residual contrasts involving internal emotionality, bodily threat, aesthetic stimulation, macro-level authority, and everyday control (Sikora et al., 27 May 2026).

6. Limitations, robustness, and methodological cautions

Across formulations, SSD assumes that the meaning–outcome relation is adequately captured by a linear direction in embedding space. Several papers state this explicitly as a limitation. Nonlinearity may require kernel methods, local models, autoencoder bottlenecks, supervised representation learning, ICA, factor analysis, or supervised PCA, but these are presented as extensions rather than part of the established core workflow (Ostrowicki et al., 26 May 2026).

A second limitation is dependence on embedding geometry and corpus composition. Published analyses repeatedly note that gradients, neighbor structure, and thematic clusters are model-dependent. The hate-speech reanalysis found that block significance was robust across embeddings while neighborhood semantics around the interaction gradient varied with the embedding geometry; the cross-lingual study reported clusters driven by toponyms, surnames, and technical vocabularies; and the shared-space paper emphasized that GloVe is context-insensitive and reflects corpus biases. This suggests that formal tests on gradients should be treated as primary inferential evidence, whereas neighborhood-based interpretation is exploratory and should be checked for stability (Sikora et al., 27 May 2026).

A third issue concerns sampling structure and measurement quality. Interaction SSD notes that coarse moderators may conflate heterogeneous subgroups and that OLS independence assumptions can be violated by repeated measures per annotator or per comment, motivating mixed-effects models or cluster-robust standard errors. The shared-space work cautions that facet-level personality analyses are underconstrained when only ten questionnaire items are available. The AI discourse study shows a related problem from the opposite direction: high-dimensional solutions can improve fit while producing semantically diffuse, weakly structured clusters (Ostrowicki et al., 26 May 2026).

Methodological best practice in this literature therefore emphasizes transparent preprocessing, explicit dimensionality sweeps, reporting of adjusted $\beta$9 and correlation or out-of-sample association measures, block tests where applicable, publication of interpretability and stability curves, and sensitivity analyses across embedding models and nearby $\hat{\beta} = (S^\top S + \lambda I)^{-1} S^\top y$0 values. Permutation tests, bootstrap intervals, and cross-validation are recurrent recommendations. The cross-lingual work additionally notes that multiple-comparison correction was not applied and advises Holm–Bonferroni or FDR control when many pairwise or dimensional tests are performed (Plisiecki et al., 13 Mar 2026).

A final interpretive caution concerns reification. Interaction SSD explicitly states that the interaction gradient is a divergence axis rather than an absolute outcome axis for either group, and the same paper warns against essentializing groups or using gradients for profiling. More broadly, SSD estimates semantics of measurement language as encoded in a particular embedding space and linked to a particular supervision signal; it does not identify a construct in isolation from wording, corpus history, or annotation design.