Comparing Human Expertise and Large Language Models Embeddings in Content Validity Assessment of Personality Tests

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Nicola Milano, Michela Ponticorvo, Davide Marocco

Keywords: Large Language Models, Personality, Big Five, Psychometrics, Persona

Source: Open primary source (opens in a new tab)

3
Authors
14
Findings
43
Limitations
11
Evidence

Editorial summary

English

This preprint compares ten graduate psychology students with five MPNet embedding models on recovery of the officially assigned Big Five factor for items from the Italian Big Five Questionnaire (BFQ) and the Big Five Inventory (BFI). For BFQ, 50 of 132 items are selected, ten per factor. The method introduces the 44-item BFI, but results, percentages, and plots use eight items per factor, 40 total, without explaining which four are excluded. Italian judges score each item against each construct as 0, 1, or 2. Although the paper calls this the Content Validity Ratio, final classification does not use the published CVR statistic or its critical threshold; it selects the construct receiving the most positive ratings. Models are paraphrase-multilingual-mpnet-base-v2 on Italian, all-mpnet-base-v2 on an English translation, SurveyBot3000, dwulff/mpnet-personality, and a new all-MPNet variant fine-tuned by the authors. For each item, the algorithm averages cosine similarity to all other items whose theoretical labels are already known, separately by factor, and selects the maximum after softmax. This is not unsupervised: it uses labels for every other benchmark item and is transductive leave-one-item-out classification. Softmax adds neither evidence nor calibrated probability; it preserves the same argmax. The useful result is descriptive. On BFQ the final table reports human 84%, multilingual MPNet 64%, English MPNet 70%, SurveyBot 64%, Personality MPNet 72%, and the authors’ model 80%. On BFI it reports human 72%, multilingual 77.5%, English 80%, SurveyBot 75%, Personality MPNet 97.5%, and the authors’ model 82.5%. Personality MPNet’s BFI advantage corresponds to recovering 39/40 labels versus 29/40 for humans; it does not establish content validity or superiority over test-construction experts. The task evaluates recovery of official labels, not whether an item set comprehensively and representatively covers a construct domain. Facet coverage, relevance, clarity, cultural bias, and construct-irrelevant variance are not measured. Language comparison is confounded: humans and multilingual MPNet see Italian, while English models see undocumented translations; no crossed design separates model, language, and translation. The new model is trained on IPIP pairs pseudo-labeled by the retrieval cross-encoder cross-encoder/ms-marco-MiniLM-L12-v2. The paper says 3,250 items yield more than 15 million pairs, but there are 5,279,625 unordered pairs and 10,559,250 ordered non-self pairs; with the 3,320 items mentioned elsewhere, counts would be 5,509,540 or 11,019,080. It also describes cross-encoder outputs as similarities from −1 to 1, while the official model card shows logits such as 9.219 and −4.078. No normalization, clipping, or loss details are given. The authors omit a train/test split because exact BFI and BFQ items are said not to occur in IPIP, but this does not evaluate fine-tuning generalization, rule out near-duplicate inventory items, or validate learning from millions of dependent pairs. Promised supplementary material is absent from the PDF and was not located elsewhere: the item list, Figures S1–S16, and training details are missing. Internal contradictions remain. Human BFQ is repeatedly described as 82%, while factor results and the final table sum to 42/50=84%. Human BFI sums to 29/40=72.5%, not 72%. Personality MPNet is said to raise BFQ Neuroticism from 50% to 100%, but base MPNet is already reported at 100%. The authors’ model is said to improve all BFQ constructs, although it ties base MPNet on Agreeableness and Neuroticism. “Significant” is used without a test, interval, or repeated sample; 80% versus 84% on BFQ is two items. The faithful conclusion is that some embeddings recover Big Five labels accurately on these two sets under a transductive procedure and that a model trained on inventory correlations reaches 39/40 on the BFI subset. The study does not automatically validate test content, demonstrate objectivity or robustness, or establish a hybrid advantage because no system combining human and model decisions is tested.

Español

Este preprint compara a diez estudiantes de posgrado en Psicología con cinco modelos de embeddings MPNet para recuperar el factor Big Five asignado oficialmente a ítems del Big Five Questionnaire italiano (BFQ) y del Big Five Inventory (BFI). Para BFQ se seleccionan 50 de 132 ítems, diez por factor; para BFI el método presenta el inventario de 44 ítems, pero resultados, porcentajes y gráficos usan ocho ítems por factor, 40 en total, sin explicar qué cuatro se excluyen. Los jueces italianos puntúan cada ítem frente a cada constructo con 0, 1 o 2. Aunque el paper llama a esto Content Validity Ratio, la clasificación final no usa el CVR publicado ni su umbral crítico: elige el constructo con más valoraciones positivas. Los modelos son paraphrase-multilingual-mpnet-base-v2 sobre italiano, all-mpnet-base-v2 sobre una traducción inglesa, SurveyBot3000, dwulff/mpnet-personality y una nueva variante de all-MPNet afinada por los autores. Para cada ítem, el algoritmo calcula su similitud coseno media con los demás ítems cuyas etiquetas teóricas ya se conocen, por factor, y elige el máximo tras softmax. Este método no es no supervisado: usa las etiquetas de todos los otros ítems del mismo benchmark y es una clasificación transductiva leave-one-item-out. Softmax no añade evidencia ni calibra probabilidades; conserva el mismo argmax. El resultado útil es descriptivo. En BFQ, la tabla final reporta humano 84%, MPNet multilingüe 64%, MPNet inglés 70%, SurveyBot 64%, Personality MPNet 72% y el modelo de los autores 80%. En BFI reporta humano 72%, multilingüe 77,5%, inglés 80%, SurveyBot 75%, Personality MPNet 97,5% y autores 82,5%. La ventaja del modelo personality sobre humanos en BFI equivale a 39/40 frente a 29/40 etiquetas recuperadas; no demuestra validez de contenido ni superioridad sobre expertos en construcción de tests. De hecho, la tarea evalúa recuperación de etiquetas oficiales, no si el conjunto cubre de forma representativa y exhaustiva el dominio del constructo. No se mide cobertura de facetas, relevancia, claridad, sesgo cultural o varianza irrelevante. La comparación de idiomas también está confundida: humanos y MPNet multilingüe ven italiano, mientras los modelos ingleses ven traducciones no documentadas; no hay un diseño cruzado que separe modelo, idioma y traducción. El nuevo modelo se entrena con pares IPIP pseudoetiquetados por el cross-encoder de ranking cross-encoder/ms-marco-MiniLM-L12-v2. El paper afirma que 3.250 ítems producen más de 15 millones de pares, pero combinaciones no ordenadas son 5.279.625 y pares ordenados sin identidad 10.559.250; con los 3.320 ítems mencionados en otro párrafo serían 5.509.540 o 11.019.080. También describe las salidas del cross-encoder como similitudes entre −1 y 1, mientras la tarjeta oficial muestra logits como 9,219 y −4,078. No explica normalización, clipping o función de pérdida. Los autores justifican no crear train/test porque BFI y BFQ exactos no aparecen en IPIP, pero eso no evalúa generalización del afinado, no evita casi duplicados entre inventarios y deja sin validación el aprendizaje sobre millones de pares dependientes. El suplemento prometido no está en el PDF ni se localizó fuera: faltan lista de ítems, gráficos S1–S16 y detalles de entrenamiento. Hay además contradicciones internas. El BFQ humano se describe repetidamente como 82%, aunque los cinco resultados por factor y la tabla suman 42/50=84%. El BFI humano suma 29/40=72,5%, no 72%. Se afirma que Personality MPNet elevó neuroticismo BFQ de 50% a 100%, pero el MPNet base ya figura con 100%. Se dice que el modelo de los autores mejora todos los constructos BFQ, aunque iguala al base en amabilidad y neuroticismo. “Significativo” se usa sin test, intervalo o repetición; 80% frente a 84% en BFQ son solo dos ítems. La conclusión fiel es que ciertos embeddings recuperan con alta exactitud las etiquetas Big Five de estos dos conjuntos bajo un procedimiento transductivo y que el modelo entrenado con correlaciones de inventarios obtiene 39/40 en el subconjunto BFI. El trabajo no valida automáticamente el contenido de un test, no demuestra objetividad ni robustez, y no establece una complementariedad híbrida porque no prueba ningún sistema que combine decisiones humanas y de modelo.

Research question

How accurately do ten judges and several MPNet embeddings recover the official Big Five label of BFQ/BFI items, how do language and fine-tuning influence this, and what are the limits of using semantic similarity as an aid to content validity?

Method

Ten Italian students score each item against five factors on a 0–2 scale and the factor with the most positive ratings is assigned. Five sentence encoders classify each item by its mean similarity with all other already-labeled items of each factor, in transductive leave-one-item-out. Total and per-factor accuracy are reported on 50 BFQ items and an implicit BFI subset of 40. An MPNet variant is fine-tuned with all IPIP pairs pseudo-labeled by an MS MARCO cross-encoder.

Sample: Panel of ten Psychology graduate students at the University of Naples Federico II, 60% women and 40% men. No psychometric experience, training, ages, inter-judge reliability, per-item CVR, agreement, consent, or compensation are reported. The automatic evaluation contains 50 BFQ and 40 BFI inferred from the tables.

Findings

  • The current source is arXiv:2503.12080v1, submitted on 15 March 2025, with 34 pages.
  • The 34 pages were rendered and visually inspected; page 34 is blank.
  • The current PDF matches byte for byte with the cache and has SHA-256 adccd496547b9527c4f504194b715111617311acede560444f3ba1b05ec4dd20.
  • BFQ: the final table reports human 84%, multilingual 64%, MPNet English 70%, SurveyBot 64%, Personality MPNet 72%, and the authors' model 80%.
  • BFI: the final table reports human 72%, multilingual 77.5%, MPNet English 80%, SurveyBot 75%, Personality MPNet 97.5%, and the authors' model 82.5%.
  • Personality MPNet recovers 39 of 40 implicit BFI labels.
  • The authors' model recovers 40 of 50 BFQ and 33 of 40 BFI.
  • The judges recover 42 of 50 BFQ according to the per-factor table and 29 of 40 BFI.
  • The automatic method uses the known labels of all remaining items to classify each item.
  • Softmax does not change which construct maximizes the mean cosine similarity.
  • 3,250 items generate 5,279,625 unordered pairs or 10,559,250 ordered pairs without identity, no more than 15 million.
  • 3,320 items would generate 5,509,540 unordered pairs or 11,019,080 ordered pairs without identity.
  • The official cross-encoder model card shows logits outside [−1,1], contradicting the declared range.
  • The promised supplement, the new checkpoint, code, or item-level results were not located.

Limitations

  • The task is label recovery of factor, not a complete evaluation of content validity.
  • Domain exhaustiveness, facet coverage, representativeness, clarity, relevance, or cultural bias are not evaluated.
  • Items are compared against official labels treated as truth, a limitation the authors partially acknowledge.
  • The judges are graduate students; no experience is accredited as a panel of specialists in the two instruments.
  • No judge training, provided definitions, order, blinding, or duration are reported.
  • No inter-judge agreement, kappa, ICC, rating distribution, or uncertainty is reported.
  • Human classification uses a count of positive ratings, not the CVR defined in the methodological section.
  • No per-item CVRs are published, nor is the critical threshold of a ten-judge panel applied.
  • The BFQ uses a stratified sample of 50 items without a seed or a reproducible list.
  • The BFI is described with 44 items but results and graphs imply 40; four are missing without explanation.
  • The automatic algorithm knows the labels of all other items in the benchmark.
  • Calling it unsupervised hides that it is a transductive form of propagation from labels.
  • There is no inductive evaluation on a completely new test or constructs without labeled anchor items.
  • The similarities of an item are averaged with groups of potentially different sizes in BFI.
  • Softmax produces numbers that sum to one but not calibrated membership probabilities.
  • No margins, entropy, top-2, calibration, or ambiguous cases are reported.
  • Only accuracy is used; there are no numerical matrices, macro-F1, balanced accuracy, or binomial intervals.
  • There are no tests of differences between humans and models or between checkpoints.
  • “Significant” is used for changes of percentage points without statistical inference.
  • The Italian-English comparison confuses architecture, training corpus, and translation.
  • It is not indicated who translated, from which version, with what quality control, or whether official English items were used.
  • Humans and English MPNet do not receive exactly the same text.
  • The best BFI model may have seen common or nearly duplicated items in training inventories; overlap is not audited.
  • The dwulff/mpnet-personality model card itself warns of high performance due to memorization on common items.
  • Excluding exact BFQ/BFI matches from IPIP does not exclude paraphrases or near duplicates.
  • The IPIP dataset is quantified inconsistently as more than 3,320 and 3,250 items.
  • The figure of more than 15 million pairs contradicts the printed combinatorial formula.
  • The pairs repeatedly share the same items and are not independent observations.
  • No training, validation, or test set is created for the authors' model.
  • The absence of exact BFQ/BFI items does not justify omitting evaluation of the fine-tuned model itself.
  • The MS MARCO cross-encoder is trained for passage ranking, not validated here as a judge of psychological similarity.
  • Its logits are not restricted to [−1,1], contrary to the methodological description.
  • It is not explained how logits are transformed to train cosine, nor loss, epochs, batch, seed, or checkpoint selection.
  • The final model is called Cross Encoder MPNet although at inference it produces bi-encoder embeddings.
  • The supplement cited for items, graphs, and training is not part of the available PDF.
  • The human BFQ appears as 82% in prose and 84% in the table and sum by constructs.
  • The human BFI appears as 72%, although the five percentages equal 29/40=72.5%.
  • The text attributes a BFQ neuroticism improvement from 50% to 100%, but the base MPNet already figures at 100%.
  • The authors' model does not improve all BFQ factors; it ties with the base on agreeableness and neuroticism.
  • Figure 4 is labeled as validator ratings although it represents the authors' model.
  • No hybrid algorithm combining judges and embeddings is tested.
  • No replication on other instruments, languages, cultures, or constructs outside Big Five is reported.
  • No code, authors' checkpoint, derived data, or peer-reviewed publication was located.

What the study does not establish

  • It does not establish complete content validity of BFQ or BFI.
  • It does not demonstrate that models discover constructs without prior labels.
  • It does not demonstrate general superiority of AI over psychometric experts.
  • It does not demonstrate that 97.5% generalizes to new items or tests.
  • It does not rule out memorization or semantic overlap with training data.
  • It does not demonstrate that English is intrinsically superior to Italian.
  • It does not demonstrate statistical significance of accuracy differences.
  • It does not demonstrate calibrated probabilities of construct membership.
  • It does not validate the MS MARCO cross-encoder as a generator of psychological similarity.
  • It does not demonstrate the effectiveness of a hybrid human-AI system because none is implemented.
  • It does not justify replacing expert panels in test development.

Traceability

Scope: Full text

Version: arXiv:2503.12080v1, submitted 15 March 2025, 34 pages

Consulted source: https://arxiv.org/abs/2503.12080

Review: Codex complete bilingual full-text fidelity pass, current arXiv-version and byte-level PDF check, all-page visual inspection, label-recovery versus content-validity audit, CVR-procedure audit, transductive-label-use audit, pair-count reconstruction, official Hugging Face teacher-output verification, table/prose reconciliation, contamination and reproducibility assessment; summaries written from the full paper and tables rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • paraphrase-multilingual-mpnet-base-v2, called MPNet Multilingual/BERTMultilingual in the paper
  • sentence-transformers/all-mpnet-base-v2 on English translations
  • SurveyBot3000 MPNet from Hommel and Arslan
  • dwulff/mpnet-personality from Wulff and Mata
  • Authors’ all-MPNet bi-encoder fine-tuned on IPIP pairs using cross-encoder/ms-marco-MiniLM-L12-v2 pseudo-labels, called Cross Encoder MPNet

Instruments and metrics

  • Big Five Questionnaire, 132-item Italian instrument with 50 stratified items used
  • Big Five Inventory, described as 44 items but 40 implicitly evaluated
  • Three-level human relevance rating labeled as Content Validity Ratio
  • Cosine similarity to construct-labeled benchmark items
  • Softmax argmax assignment
  • Overall and construct-level label-recovery accuracy

Data used

  • 50 BFQ items: ten sampled items for each Big Five factor; exact list and random seed unavailable
  • 40 implied BFI items: eight per factor; selection from BFI-44 unexplained
  • International Personality Item Pool: stated as over 3,320 items and later as 3,250 items
  • Pseudo-labels from cross-encoder/ms-marco-MiniLM-L12-v2
  • No released item table, supplementary figures, training code, trained authors’ checkpoint or prediction matrix located

Evidence and location

  • Version, title, authors, and abstract: arXiv:2503.12080v1 metadata and pages 1–2 checked 15 July 2026
  • Instruments, sampling, and human panel: Sections 2.1–2.2, pages 7–9
  • Base models and fine-tuning process: Sections 2.3–2.4, pages 9–14
  • Transductive classification algorithm: Section 2.5, page 15
  • BFQ results: Section 3.1, pages 16–21; final table, page 26
  • BFI results: Section 3.2, pages 21–27; final table, pages 26–27
  • Limitations and authors' conclusions: Discussion and Conclusion, pages 27–31
  • Real range of the MS MARCO cross-encoder: Official Hugging Face model card checked 15 July 2026; example scores 9.218911 and −4.0780287
  • Personality MPNet memorization risk: Official dwulff/mpnet-personality Hugging Face model card checked 15 July 2026
  • Arithmetic, inconsistencies, and artifact status: reports/verification/article-184-method-and-arithmetic-audit.json
  • Complete visual inspection: All 34 PDF pages rendered and visually inspected on 15 July 2026