Revised Portrait Values Questionnaire (PVQ-RR)

• PVQ-RR is a psychometric tool that measures ten basic human values through 57 items organized into 19 facets based on Schwartz’s circumplex model.
• It leverages the SQuID embedding methodology to recover latent value structures by centering neural representations and analyzing cosine similarities.
• The tool demonstrates high internal consistency and strong alignment with human-derived data, providing an efficient, scalable approach to measure validation.

The Revised Portrait Values Questionnaire (PVQ-RR) is a psychometric assessment instrument designed to measure human value dimensions as defined by Schwartz’s theory. The PVQ-RR comprises 57 succinct “portrait” statements, each prompting respondents to rate the extent of self-description on a six-point Likert scale ranging from “not like me at all” to “very much like me.” These 57 items are systematically organized into 19 narrowly defined facets, each comprising three items, which further aggregate into ten basic human values: Self-Direction, Stimulation, Hedonism, Achievement, Power, Security, Conformity, Tradition, Benevolence, and Universalism. Schwartz's circumplex model positions these values in a circular arrangement, such that proximate values possess compatible motivations and predominantly positive correlations, while values on opposite sides, such as Self-Direction versus Conformity, exhibit negative correlations. Multidimensional scaling (MDS) of inter-value correlation matrices is customarily employed to reveal this circular order and polarity structure (Pellert et al., 29 Sep 2025).

1. PVQ-RR Structure and Theoretical Basis

The PVQ-RR's construction reflects the goal of capturing detailed aspects of value orientations. Each of its 57 items describes behavioral goals or guiding life principles attributed to a hypothetical person and asks respondents to indicate their similarity to that portrait. The aggregation of three items per facet yields 19 distinct subdimensions, mapped onto the ten higher-order values postulated by Schwartz’s theory. The theoretical circumplex arrangement derived from Schwartz’s model provides an explicit geometric representation of the motivational continuities and oppositions among value dimensions—quantitatively validated through inter-correlation patterns and their spatial unfolding via MDS.

2. Embedding-Based Recovery of PVQ-RR Structure via SQuID

The Survey and Questionnaire Item Embeddings Differentials (SQuID) methodology enables the extraction of latent value structures directly from the PVQ-RR items using general-purpose neural network embeddings, obviating the requirement for extensive human rater data. The key steps are as follows:

1. Item Embedding Acquisition: Each PVQ-RR item, averaging across pronoun variants, is passed through off-the-shelf embedding models (e.g., Linq-Embed-Mistral, gemini-embedding-exp-03-07, jina-embeddings-v3, KaLM-embedding-multilingual-mini-instruct-v1.5, mpnet-personality) to generate high-dimensional vector representations.
2. Questionnaire-Level Subtraction: To address the conflation of item-specific semantics and shared linguistic background (which hinders the emergence of negative correlations), SQuID subtracts the mean embedding of all questionnaire items from each item's embedding:

$$\bar e = \frac1{57}\sum_{i=1}^{57} e_i,\quad \tilde e_i = e_i - \bar e.$$

This centering operation is empirically sufficient for recovering both positive and negative inter-item correlations in the absence of domain-specific fine-tuning.

1. Construction of Value-Dimension Vectors: For each value dimension $d$, the set of positively keyed items $I^+_d$ is averaged post-centering to yield a "directional differential" vector $\Delta_d$. With PVQ-RR’s solely forward-keyed structure, reverse-keyed terms are omitted:

$$\Delta_d = \frac{1}{|I^+_d|} \sum_{i\in I^+_d} \tilde e_i.$$

Optional $\ell_2$ normalization of $\Delta_d$ facilitates focus on angular (cosine-based) inter-value relationships.

3. Quantitative Analysis and Evaluation Metrics

The alignment between embedding-derived and human-derived latent value structures is quantitatively assessed using multiple classical and embedding-based psychometric metrics:

Dimension–Dimension Similarity:

• Cosine similarity between value vectors in the embedding space:

$$S_{dd'}^{(\mathrm{emb})} = \cos(\Delta_d, \Delta_{d'}) = \frac{\Delta_d^\top \Delta_{d'}}{\|\Delta_d\|\;\|\Delta_{d'}\|}.$$

• Human data yields $$S_{dd'}^{(h)} = \mathrm{corr}(v_d^{(h)}, v_{d'}^{(h)})$$ via Pearson correlation.

Statistical Correspondence:

• Linear regression relates $S_{dd'}^{(\mathrm{emb})}$ to $d$0, with coefficient of determination $d$1 reflecting explained variance. SQuID achieves $d$2 with Linq-Embed-Mistral, signifying 55% of variance in human dimension-dimension similarities is captured without task-specific fine-tuning.

Internal Consistency (Cronbach’s α):

• Cronbach’s $d$3 computed on embedding coordinates as analogues of rater responses yields $d$4 for embedding-derived facets, compared to the human-data average of 0.70.

Factor Congruence:

• Tucker congruence coefficients between 2-D MDS solutions derived from human and embedding data reach values of 0.88 and 0.82 along principal axes, classed as "fair" by conventional criteria.

Multidimensional Scaling (MDS):

• Dissimilarity matrices $d$5 are projected via MDS into a two-dimensional space, faithfully reproducing the theoretical circumplex structure where values are sequenced in a circular topology.

Procrustes Alignment:

• Embedding-derived and human-derived MDS configurations are aligned using orthogonal Procrustes analysis to optimize congruence and visual correspondence.

4. Summary of Empirical Findings

Evaluation of SQuID on PVQ-RR yields several notable results:

Metric	Embedding Model (Best)	Human Data
Explained variance ($d$6)	0.55	–
Cronbach's $d$7	0.77 (Linq-Embed-Mistral)	0.70
Tucker congruence	0.88 / 0.82 (axes 1/2)	N/A

• SQuID recovers negative and positive inter-value correlations characteristic of opposing and aligned motivational types, respectively.
• The embedding-based MDS solution produces a structurally identifiable circumplex with the appropriate ordering and separation of the ten fundamental value dimensions.
• The best-performing embedding model closely tracks the empirical inter-value correlation structure, surpassing human-data benchmarks in internal consistency per facet (Pellert et al., 29 Sep 2025).

5. Methodological Advantages and Constraints

SQuID demonstrates several methodological properties:

• Efficiency and Scalability: The procedure requires only a single pass of PVQ-RR items through established embedding models, followed by a computationally trivial mean-subtraction. This sharply reduces data collection time, circumvents issues of rater availability, fatigue, or literacy, and is practical on commodity hardware.
• Flexibility: SQuID's approach is agnostic to embedding model choice and accommodates arbitrary questionnaire structures without retraining or fine-tuning, preserving the ability to capture negative correlations central to value theory.
• Psychometric Complementarity: Embedding-derived results can provide rapid “in silico pre-tests” during measure development phases, identify problematic items, or reveal latent structures prior to field deployment with human subjects.
• Limitations: Subsequent investigations must clarify the impact of pretrained model memorization of widely distributed questionnaires, techniques for appropriately reverse-keying item embeddings, and the generalizability to multilingual and cross-cultural contexts. It is advised that embedding-based “norms” supplement rather than replace empirical human data to avoid misrepresenting population variability through the filter of LLM knowledge.

6. Broader Implications and Extensions

The SQuID methodology’s performance on PVQ-RR illustrates its potential as a complementary analytic tool for psychometrics and the behavioral sciences. It enables rapid, reproducible mapping of complex latent construct topologies from survey text alone. There is scope for application beyond human values, including attitudinal, personality, or ideological domains. A plausible implication is that such embedding-based approaches may streamline initial measure validation workflows and extend psychometric analyses into areas with limited access to large, diverse respondent pools. Caution is warranted to maintain human-grounded validity standards, especially in the establishment of population norms and the detection of construct variance beyond LLM priors (Pellert et al., 29 Sep 2025).