Revised Portrait Value Questionnaire (PVQ-RR)

• Revised Portrait Value Questionnaire (PVQ-RR) is a psychometric instrument assessing 19 fine-grained human values using 57 items (3 per value) based on Schwartz’s theory.
• It employs a 6-point Likert scale and averaging approach, enabling structured analysis of motivational profiles and circumplex relationships among values.
• Computational embedding methods, augmented by the SQuID procedure, replicate human-derived metrics with strong internal consistency and significant global validation.

The Revised Portrait Value Questionnaire (PVQ-RR) is a psychometric instrument designed to measure 19 fine-grained value dimensions within Schwartz’s theory of basic human values. Each dimension is assessed via three items, each formatted as a short statement (e.g., "It is important to him/her to …") and rated on a 6-point Likert scale ranging from "not like me at all" to "very much like me." Dimension scores are calculated by straightforward averaging across the three relevant items, allowing for structured assessment of an individual’s motivational value profile. PVQ-RR’s structure explicitly supports analyses of multidimensional circumplex relationships among human values and is widely validated via large international human datasets (Pellert et al., 29 Sep 2025).

1. Instrument Structure and Scoring

The PVQ-RR comprises 57 items mapped a priori into 19 basic values, such as "Self-direction–Thought," "Security–Societal," and "Hedonism," with each basic value associated with exactly three items. The notation $I_n$ denotes the index set of items for value $n$ ($|I_n|=3$). Scores for each dimension $V_n$ are determined by:

$V_n = \frac{1}{|I_n|} \sum_{i \in I_n} r_i$

where $r_i$ is the respondent’s Likert rating (or, in computational analogues, the derived score from item embeddings).

This organization facilitates explicit mapping between survey items and the underlying value construct, providing both theoretical transparency and analytic tractability for dimension-level and circumplex-level investigations.

2. Embedding Extraction from PVQ-RR Items

To operationalize the PVQ-RR in a computational setting, off-the-shelf LLM sentence-embedding architectures are employed. Five pre-trained models have been evaluated:

Model Name	Parameters	Notable Feature
Linq-Embed-Mistral	7.11B	Top performer in MTEB STS
gemini-embedding-exp-03-07	Not disclosed	Google, closed-API
jina-embeddings-v3	572M	General-purpose
KaLM-embedding-multilingual-mini-instruct-v1.5	494M	Multilingual capability
mpnet-personality	109M	Finetuned for personality item correlations

Each PVQ-RR item is encoded as a fixed-length embedding $y_i \in \mathbb{R}^n$ (dimensionality dependent on model). To mitigate gender bias, items are prompted with both male- and female-pronoun versions, with embeddings averaged when supported.

3. Survey and Questionnaire Item Embeddings Differentials (SQuID)

Raw sentence embeddings for PVQ-RR items are typically highly similar, yielding predominantly nonnegative inter-item correlations. To uncover negative relationships essential to the theoretical value circumplex, the Survey and Questionnaire Item Embeddings Differentials (SQuID) methodology is applied:

• Compute the questionnaire centroid: $\bar{y} = \frac{1}{57} \sum_{i=1}^{57} y_i$.
• Center each item: $\delta_i = y_i - \bar{y}$.
• Aggregate within each value: $e_n = \frac{1}{|I_n|} \sum_{i \in I_n} \delta_i$.

This centering operation eliminates global linguistic biases, revealing both positive and negative inter-dimension relationships in accordance with Schwartz's circumplex theory. The approach is analogous to vector arithmetic in distributional semantics, and it enables the recovery of structure needed to differentiate opposing value dimensions.

4. Evaluation Metrics

The alignment between embedding-derived PVQ-RR scores and human data is quantified along several axes:

• Internal Consistency (Cronbach’s $n$0):

$n$1

Applied to both human and embedding-derived scores.

• Dimension–Dimension Similarity:

Compute Pearson correlation matrices $n$2 (human) and $n$3 (embedding); overall similarity is:

$n$4

Variance explained: $n$5 from regression of $n$6 on $n$7.

• Factor-Congruence Coefficient (for MDS axes):

$n$8

Where $n$9 and $|I_n|=3$0 are 2D MDS coordinates after Procrustes alignment.

• Recovery of Negative Correlations:

Assessed by the sign and structure of off-diagonal entries in the SQuID-corrected item–item similarity matrix (substantial negatives required for circumplex compatibility).

All main results are evaluated for significance via Monte Carlo shuffling ($|I_n|=3$1).

5. Multidimensional Scaling and Structure Recovery

The SQuID-processed embedding-derived correlation matrices are subjected to ordinal (nonmetric) MDS (as implemented in the R package smacof), with dissimilarities defined as $|I_n|=3$2. Both embedding and human-derived MDS solutions exhibit the hallmark Schwartz circumplex: motivationally adjacent values cluster, and opposed values (e.g., "Conformity" vs. "Self-direction") are maximally separated across the circumplex. Procrustes matching aligns embedding and human MDS solutions, facilitating the quantification of structural congruence (factor-congruence coefficients $|I_n|=3$3, $|I_n|=3$4).

6. Quantitative Performance and Empirical Benchmarks

The embedding-based approach, after SQuID correction, achieves results commensurate with large human datasets on key benchmarks:

• Internal Consistency:

Average Cronbach’s $|I_n|=3$5 (Linq-Embed-Mistral) versus human benchmark $|I_n|=3$6; random embedding baseline $|I_n|=3$7.

• Dimension–Dimension Correlation:

Pearson $|I_n|=3$8 (95% CI [0.66, 0.80]); $|I_n|=3$9 variance explained.

• Factor Congruence (MDS Axes):

$V_n$0, $V_n$1. Both indices exceed the conventional "fair" threshold ($V_n$2).

• Statistical Significance:

All core benchmarks are significant at $V_n$3.

These findings establish that SQuID-processed sentence embeddings reproduce key psychometric structure—from item group interrelations to the full circumplex arrangement—at levels on par with survey data (Pellert et al., 29 Sep 2025).

7. Theoretical Implications, Applications, and Limitations

Embedding-based recovery of PVQ-RR structure demonstrates that LLMs encode not only semantic but also fine-grained motivational distinctions postulated by Schwartz’s value theory. This suggests that LLMs’ text-derived internal geometries are congruent with those revealed by extensive psychometric fieldwork. Derived benefits include cost-free in-silico pretesting of new items or scales, rapid scale revision, and global language/culture coverage without additional data collection.

Potential limitations arise from:

• Possible memorization of the PVQ-RR within embedding model training corpora.
• Open questions regarding the treatment of reverse-keyed items in embedding space.
• General inferential risks in analogizing embedding-derived similarities to human judgments.

A plausible implication is that neural embedding techniques, when combined with minimal centering corrections, provide a complementary methodology for large-scale psychometric scale development and validation, expanding the scope and flexibility of behavioral measurement paradigms (Pellert et al., 29 Sep 2025).