LMLPA: Language Model Linguistic Personality Assessment

Evaluation and psychometric validity2024ACL AnthologyApproved editorial review

Authors: Jingyao Zheng, Xian Wang, Simo Hosio, Xiaoxian Xu, Lik-Hang Lee

Keywords: Computation and Language, Artificial Intelligence, Personality Assessment, Computational Linguistics

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

5
Authors
16
Findings
34
Limitations
15
Evidence

Editorial summary

English

LMLPA proposes replacing direct administration of the Big Five Inventory to an LLM with a questionnaire adapted to linguistic capabilities. It converts the 44 BFI items into open-ended “to what extent” questions, requires the model to include always, often, sometimes, rarely, or never plus a one-sentence rationale, and asks another model to turn the text into a 1–5 score. Five psychology professionals qualitatively reviewed the adaptations; PCA then removes four items, leaving a final 40-item version. Its clearest contribution is the complete publication of questions, prompts, labels, and order-sensitivity analyses, while explicitly acknowledging that LLMs do not have human actions, emotions, or cognitive processes and that output is used as a linguistic proxy.

In the order tests, GPT-4-Turbo answers all 44 items in separate calls at temperature 0. Reversing the required adverb list yields seven pairs considered truly inconsistent after one semantically equivalent case is manually corrected, with weighted kappa 0.730. In the later comparison, numeric BFI produces 16 discrepancies and kappa 0.401, whereas the full LMLPA system produces six and kappa 0.877. This demonstrates greater output stability under those two configurations, but it does not isolate the effect of an open questionnaire: the BFI condition reverses the scale shown to the test-taking model, while LMLPA reverses the scale shown to GPT-4-Turbo as judge. An ordered list of five adverbs remains in the test-taker prompt, and a list of five labels is shifted to the rater, so option structure is moved to another stage rather than removed.

Three of the experts rate one set of 44 GPT-4-Turbo responses. GPT-4-Turbo as rater correlates 0.827–0.877 with them and has a single-measure ICC of 0.829; BART-Large-MNLI reaches 0.700–0.810 and ICC 0.766, and Llama3-8B reaches 0.750–0.840 and ICC 0.785. This is evidence of agreement on highly structured responses containing a mandatory frequency adverb, not evidence that the inferred trait is correct. The exact ICC model and whether it targets agreement or consistency are not specified; the “average measures” ICC combines four ratings and is not the performance of one automated judge. Llama3 also obtains kappa 0.279 when its rating options are reversed, so it is excluded from the precise psychometric phase.

Reliability and structure are estimated from 250 conditions generated by the same GPT-4-Turbo: ten PersonaChat descriptions crossed with 25 profiles that explicitly place five levels of each trait in the system prompt. Alpha values on the 44 items are 0.869–0.936, but can reflect item redundancy, prompt compliance, and shared test-taker/rater method; they are not recomputed after Q31, Q35, Q41, and Q22 are removed to form the final 40-item scale. PCA reports KMO 0.951 and extracts four components rather than five; Conscientiousness is distributed across several components, cross-loadings are extensive, and the prose assigns Neuroticism to component 3 even though the table concentrates it in component 2. The same sample is used for item deletion and the validity claim, with no CFA or independent replication.

The final test again treats Big Five adjectives written directly into prompts for GPT-4-Turbo, Llama3-8B-Instruct, and Mistral-7B-Instruct-v0.2 as “ground truth.” Score distributions usually shift in the requested direction but do not match target levels; low Agreeableness clusters at 2–3, and Llama3 Neuroticism moves in the opposite direction at the extreme. No error, correlation, or classification metric is reported. When GPT-4-Turbo both generates and judges outputs, model circularity and shared instrument knowledge remain. Without exact service snapshots, seeds, repetitions, outputs, data, or public code, the full protocol is not reproducible. LMLPA is therefore a promising open instrument for describing prompt-conditioned linguistic self-reports, but its evidence does not establish a latent, stable, predictive, or human-equivalent Big Five personality.

Español

LMLPA propone sustituir la administración directa del Big Five Inventory a un LLM por un cuestionario adaptado a capacidades lingüísticas. Convierte los 44 ítems BFI en preguntas abiertas del tipo «¿hasta qué punto…?», obliga al modelo a incluir siempre, often, sometimes, rarely o never y una justificación de una frase, y encarga a otro modelo transformar el texto en una puntuación 1–5. Cinco profesionales de psicología revisaron cualitativamente las adaptaciones; después, PCA elimina cuatro ítems y deja una versión final de 40. La aportación más útil es publicar íntegramente preguntas, prompts, etiquetas y análisis de sensibilidad al orden, reconociendo además que los LLM no poseen acciones, emociones o procesos cognitivos humanos y que su output se usa como proxy lingüístico.

En las pruebas de orden, GPT-4-Turbo responde los 44 ítems en llamadas separadas y a temperatura 0. Al invertir la lista de adverbios obligatorios, siete pares se consideran realmente discordantes después de corregir manualmente un caso semánticamente equivalente, con kappa ponderado 0,730. En la comparación posterior, el BFI numérico produce 16 discrepancias y kappa 0,401, mientras el sistema completo LMLPA produce seis y kappa 0,877. Esto demuestra mayor estabilidad de outputs bajo esas dos configuraciones, pero no aísla el efecto del cuestionario abierto: en BFI se invierte la escala que ve el modelo evaluado; en LMLPA se invierte la escala que ve GPT-4-Turbo como juez. La lista ordenada de cinco adverbios sigue dentro del prompt del test-taker y la lista de cinco etiquetas se desplaza al rater, por lo que no desaparece la estructura de opción; cambia de etapa.

Tres de los expertos puntúan un único conjunto de 44 respuestas de GPT-4-Turbo. Las correlaciones del rater GPT-4-Turbo con ellos son 0,827–0,877 y su ICC de medida individual es 0,829; BART-Large-MNLI obtiene 0,700–0,810 e ICC 0,766 y Llama3-8B 0,750–0,840 e ICC 0,785. Es evidencia de acuerdo para respuestas muy estructuradas que contienen un adverbio de frecuencia obligatorio, no de exactitud del rasgo. El tipo exacto de ICC y su supuesto de acuerdo/consistencia no se especifican; el ICC «average measures» combina cuatro ratings y no representa el rendimiento de un juez individual. Llama3 muestra además kappa 0,279 cuando se invierten sus opciones, por lo que se excluye de la fase psicométrica precisa.

La consistencia y la estructura se estiman con 250 condiciones generadas por el mismo GPT-4-Turbo: diez descripciones PersonaChat combinadas con 25 perfiles que escriben explícitamente cinco niveles de cada rasgo en el system prompt. Los alfas sobre los 44 ítems son 0,869–0,936, pero pueden reflejar redundancia de ítems, obediencia al prompt, método y juez comunes; no se recalculan tras eliminar Q31, Q35, Q41 y Q22 para formar la escala final de 40. La PCA obtiene KMO 0,951 y cuatro componentes, no cinco; Conscientiousness se reparte entre varios, hay numerosas cargas cruzadas y el texto llega a atribuir Neuroticism al componente 3 aunque la tabla lo concentra en el 2. La misma muestra se usa para eliminar ítems y declarar validez, sin CFA o réplica independiente.

La prueba final vuelve a tratar como «ground truth» los adjetivos Big Five insertados directamente en el prompt de GPT-4-Turbo, Llama3-8B-Instruct y Mistral-7B-Instruct-v0.2. Las distribuciones suelen moverse en la dirección pedida, pero no coinciden con los niveles, baja agradabilidad se concentra en 2–3 y el neuroticismo de Llama3 se mueve en dirección contraria en el extremo; no se publica una métrica de error, correlación o clasificación. Cuando GPT-4-Turbo genera y juzga, existe circularidad de modelo y conocimiento compartido del instrumento. Sin snapshots precisos del servicio, semillas, repeticiones, outputs, datos o código público, el protocolo no es reproducible íntegramente. En consecuencia, LMLPA es un instrumento abierto y prometedor para describir autodeclaraciones lingüísticas condicionadas por prompts, pero sus datos no prueban una personalidad latente, estable, predictiva o equivalente al Big Five humano.

Research question

Can a BFI adapted to linguistic properties, answered via open text and scored by an AI agent, measure the "linguistic personality" expressed by LLMs more stably and psychometrically defensibly than direct administration of human scales?

Method

Development and evaluation of an instrument in several phases. It reformulates the 44 BFI items to ask about the model's linguistic behaviors and adds an instruction that requires a frequency adverb and a justification. Five psychology professionals participate in one-hour interviews to review content. GPT-4-Turbo responds and is subjected to order inversions; BART-Large-MNLI, GPT-4-Turbo, and Llama3-8B-Instruct score 44 responses and are compared with three experts via correlations and ICC. For alpha and PCA, GPT-4-Turbo generates responses under 250 combinations of ten persona descriptions and 25 Big Five profiles; PCA with Varimax eliminates four items. Finally, GPT-4-Turbo, Llama3-8B-Instruct, and Mistral-7B-Instruct-v0.2 receive five levels of each trait and are evaluated by GPT-4-Turbo and Llama3. The editorial review read and rendered the 42 pages, examined tables, figures, the 44 original items, the 40 final items, prompts, and appendices, and searched for associated public materials.

Sample: Human participation comprises five psychology professionals with 3.5 to 11 years of experience; three with self-rated Big Five familiarity of 5 re-evaluate 44 responses. No age, gender, exclusion criteria, or systematic analysis of the interviews are reported. The psychometric sample is not 250 independent LLMs: it is 250 conditions of the same GPT-4-Turbo created with ten PersonaChat texts and 25 trait instructions. In the final test, three models receive five levels for each of the five traits, each level combined with ten descriptions; the appendices show distinct persona sets for GPT-4-Turbo, Llama3, and Mistral. Calls are made at temperature 0, but exact snapshots of GPT-4-Turbo, seeds, temporal repetition, or an unambiguous total number of outputs and calls are not published.

Findings

  • The inversion of the order of adverbs in the open responses of GPT-4-Turbo produces seven real discordances out of 44 after manually correcting one semantically equivalent case; weighted kappa 0.730, 95% CI 0.554 to 0.905.
  • In the comparison of the full system, direct BFI presents 16 discrepancies and kappa 0.401, CI 0.175 to 0.626, versus six and kappa 0.877, CI 0.777 to 0.977, for LMLPA.
  • Greater stability does not eliminate ordered options: the evaluated model receives five adverbs in the prompt and the judge receives five numbered labels.
  • Responses before and after inverting adverbs have a median semantic similarity above 0.96, although MiniLM can also score high paraphrases that change the relevant frequency or polarity.
  • GPT-4-Turbo as judge correlates 0.851, 0.877, and 0.827 with the three humans; Llama3 0.80, 0.84, and 0.75; BART 0.74, 0.81, and 0.70.
  • Individual-measure ICCs are 0.829 for GPT-4-Turbo, 0.785 for Llama3, and 0.766 for BART; average-measure ICCs are 0.951, 0.936, and 0.929, respectively.
  • Llama3 obtains kappa 0.279, CI 0.109 to 0.449, when inverting the rater scale, so the authors exclude it from alpha and PCA.
  • Alphas on the 44-item version are 0.869 in extraversion, 0.899 in agreeableness, 0.924 in conscientiousness, 0.886 in neuroticism, and 0.936 in openness.
  • Q31, Q35, and Q41 do not load at least 0.40 and are eliminated over two iterations; Q22 is eliminated afterward for loading less than 0.40 on the agreeableness component, leaving 40 items.
  • Alpha is not recalculated for the final 40-item scale, despite Q35 and Q41 appearing as items whose elimination would increase the openness alpha.
  • KMO is reported as 0.951 and the PCA retains four components, not the expected five traits.
  • Openness and extraversion form relatively clear groups; agreeableness is concentrated mainly on component 3, neuroticism on 2, and conscientiousness is distributed across components 1, 2, and 3.
  • The text erroneously states that neuroticism loads predominantly on component 3, while Table 4 shows strong and negative loadings on component 2.
  • In the manipulation test, scores tend to shift in the direction of the prompt adjectives, but the observed levels do not match the target and no aggregate fidelity metric is offered.
  • The three models avoid extremely low agreeableness; Llama3's neuroticism decreases toward approximately 2 when the target prompt reaches 5.
  • Results are more concentrated when GPT-4-Turbo also acts as test-taker or rater, a pattern compatible with either better judgment or family/model circularity.

Limitations

  • The questions still request self-descriptions from the model itself, "to what extent do you do X?", so they replace self-report of emotions with self-report of capabilities and style, not with independently observed behavior.
  • The items require one of five ordered adverbs. The format is not fully open and can be mapped almost directly to the 1 to 5 scale.
  • The judge knows in advance the trait to be scored and receives explicit labels, which favors semantic recognition and reduces criterion independence.
  • Several questions conflate traits with quality, safety, or capability: factuality and coherence are assigned to conscientiousness/neuroticism, rejection of harmful language to agreeableness, and artistic knowledge to openness.
  • Some items require introspection that the model does not possess, such as knowing whether it uses its dataset efficiently or expands responses beyond its training data.
  • The interviews of five experts are qualitative; no content validity index, blind item-level ratings, analysis protocol, transcripts, or agreement on modifications are published.
  • Two experts self-assess their Big Five familiarity as 2 and 3. Experience helps design, but does not by itself turn adaptations into valid measures of the construct.
  • The declared compensation is HK$13.30 for an interview of approximately one hour; the article does not explain this unusually low figure.
  • The three humans score a single set of 44 GPT-4-Turbo responses, too narrow to validate judges across models, prompts, domains, and diverse styles.
  • Responses necessarily contain always/often/sometimes/rarely/never; humans and models can achieve high agreement by reading this surface signal without evaluating the reasoning or the construct.
  • The correlations labeled inter-item are correlations between rater vectors. No intervals, ordinal robustness, weighting, or adjustment for multiple correlations are reported.
  • The exact ICC model, selection between absolute agreement and consistency, fixed/random effects, and generalization unit are not specified, limiting interpretation and reproduction.
  • The average measures ICC is the reliability of a mean of four ratings, not evidence that the individual AI rater achieves that value.
  • The BFI to LMLPA comparison inverts scales at different stages and uses different instruments; it does not causally attribute the kappa difference to the open-ended character of the questionnaire.
  • The Q35 case is manually recoded after observing that never and always express the same meaning with opposite phrase polarity; no prior rule for similar corrections is published.
  • Kappa p<0.001 mainly tests agreement above zero; it does not demonstrate impartiality, validity, or equivalence between scales. Nor is the difference 0.401 versus 0.877 formally tested.
  • Temperature 0 does not guarantee exact determinism of hosted services, and no replications, test-retest across dates, or snapshot change analyses are performed.
  • The 250 rows come from a single GPT-4-Turbo and from prompts constructed to produce Big Five covariance; calling them "distinct GPT-4-Turbos" overestimates independence and diversity.
  • The profiles directly write the trait adjectives in the system prompt, so high alpha and factorial structure may reflect common obedience to the prompt and not a valid scale.
  • A single GPT-4-Turbo generates responses and scores them in the psychometric phase, introducing common method, model self-consistency, BFI knowledge contamination, and circularity.
  • High Cronbach alpha does not prove unidimensionality, temporal stability, measurement equivalence, or validity; it also increases with semantic redundancy among items.
  • The alphas belong to the 44 items prior to refinement; the reliability of the final 40-item version proposed for use is not documented.
  • PCA is used as factor analysis, employs orthogonal Varimax despite overlapping traits/loadings, and selects factors via Kaiser plus scree without reporting explained variance or parallel analysis.
  • KMO is accompanied by p<0.001 although KMO has no such test; Bartlett's statistic and degrees of freedom are not reported, so the attribution of p is ambiguous.
  • The structure does not recover five dimensions, conscientiousness lacks its own component, and there are numerous cross-loadings. Interpreting four components as Big Five confirmation is stronger than the data.
  • Items are eliminated and validity is declared using the same sample; there is no CFA, holdout, replication in another model, or comparison with alternative structures.
  • The text contains a material inconsistency in assigning neuroticism to component 3 when the table places it on 2.
  • The final test defines a prompt instruction as ground truth; this validates sensitivity to prompt words, not latent personality or stable behavior.
  • No MAE, target-score correlation, ordinal accuracy, intervals, or trend tests for Figure 6 are published; some conditions visibly fail and precise evaluation is nonetheless claimed.
  • The ten personas are assumed innocuous without testing and are distinct for each model, which confounds comparison of GPT-4-Turbo, Llama3, and Mistral.
  • Neutral, adversarial, or paraphrased prompts that hide trait labels are not evaluated, nor is generalization to real longitudinal conversation.
  • The article does not fix a precise GPT-4-Turbo snapshot, seeds, full hardware/configuration of open models, call dates, or number of repetitions.
  • No official repository with code, outputs, ratings, intermediate data, or scripts was identified; the appendices allow partial reimplementation, not exact reproduction of results.
  • Cultural or demographic biases, invariance, inference safety, consequences of labeling systems or humans, and predictive validity regarding interaction outcomes are not examined.

What the study does not establish

  • It does not demonstrate that GPT-4-Turbo, Llama3, or Mistral possess an internal, conscious psychological personality equivalent to that of humans.
  • It does not demonstrate that the scores are stable traits across prompts, users, tasks, languages, dates, or model versions.
  • It does not demonstrate that answering with always or often describes the model's real behavior in unguided interactions.
  • It does not demonstrate that the GPT-4-Turbo judge is objective or independent; it shows agreement with three experts on 44 highly structured responses.
  • It does not demonstrate that the kappa improvement comes exclusively from the open format, because different stages and instruments are compared.
  • It does not demonstrate Big Five validity of the final 40-item version; alpha corresponds to 44 and PCA recovers four components.
  • It does not demonstrate convergent, discriminant, criterion, predictive, ecological, or measurement invariance validity.
  • It does not demonstrate that prompt adjectives are the ground truth of personality; they are known instruction conditions for systems trained to follow natural language.
  • It does not demonstrate precise measurement of the five levels: scores do not match exactly, there is compression of extremes and directional failures.
  • It does not allow concluding that differences between models are architectural or due to safeguards, because personas, judges, and snapshots are not controlled equivalently.
  • It does not validate uses in education, marketing, manufacturing, recommendation, selection, diagnosis, or high-impact decisions.
  • It does not allow full reproduction of figures and figures without outputs, code, configuration, and analysis artifacts.
  • It does not establish a comparative benchmark of base models under neutral conditions; it mainly validates response to explicit Big Five profiles.
  • It does not prove that four components are an alternative theory of linguistic personality; it is an exploratory solution from an induced sample.

Traceability

Scope: Full text

Version: Computational Linguistics 51(2), June 2025, pp. 599–640, DOI 10.1162/coli_a_00550; final 42-page journal version with all appendices

Consulted source: https://aclanthology.org/2025.cl-2.6.pdf

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-4-Turbo as test taker and AI rater
  • Meta-Llama-3-8B-Instruct as test taker and AI rater
  • Mistral-7B-Instruct-v0.2 as test taker
  • facebook/bart-large-mnli as zero-shot NLI rater
  • sentence-transformers/all-MiniLM-L6-v2 for response similarity

Instruments and metrics

  • Original 44-item Big Five Inventory
  • 44-item Adapted-BFI open-ended draft
  • Final 40-item LMLPA questionnaire after PCA
  • Five mandatory frequency adverbs: always, often, sometimes, rarely, never
  • GPT rater 1–5 Big Five label prompt
  • BART-Large-MNLI entailment-based zero-shot classifier
  • Weighted Cohen kappa
  • Sentence embedding cosine similarity
  • Pairwise rater correlation
  • Intraclass correlation coefficient
  • Cronbach alpha
  • Kaiser-Meyer-Olkin measure
  • Bartlett test of sphericity
  • Principal component analysis with Varimax rotation

Data used

  • One 44-response GPT-4-Turbo set for order and rater-agreement experiments
  • Five psychology expert interviews
  • Three expert ratings of the same 44 GPT-4-Turbo responses
  • 250 GPT-4-Turbo prompt conditions: 10 PersonaChat descriptions × 25 Big Five profiles
  • PersonaChat persona descriptions from Zhang et al. (2018)
  • Five prompted levels per trait and 10 persona descriptions per test-taker model in the final measurement test
  • Complete questionnaire, prompts, labels, persona descriptions and loadings published in the article appendices
  • No public raw outputs, analysis code or machine-readable result dataset identified

Evidence and location

  • Final publication, DOI, scope, and acknowledgment of human to LLM differences: Computational Linguistics 51(2), pp. 599 to 602, abstract and section 1
  • BFI adaptation and review by five experts: Published article, sections 3.1.1 to 3.1.2 and Table 2, pp. 605 to 608
  • Mandatory adverbs, justification, and first inversion experiment: Published article, sections 3.1.3 to 3.2 and Figure 2, pp. 608 to 610
  • BART, GPT-4-Turbo, Llama3, and comparison with three humans: Published article, sections 4.1 to 4.2 and Figure 3, pp. 611 to 614
  • ICCs, Llama3 kappa, and selection of GPT-4-Turbo as psychometric judge: Published article, section 4.2.2 and opening of section 5, pp. 613 to 614
  • BFI versus LMLPA under scale inversion: Published article, section 5.1 and Figure 4, pp. 615 to 616
  • 250 GPT-4-Turbo conditions constructed with ten personas and 25 profiles: Published article, section 5.2, p. 617; Appendix E, pp. 631 to 633
  • Alphas on 44 items: Published article, section 5.2.1 and Table 3, pp. 617 to 618
  • KMO, PCA, four components, eliminations, and cross-loadings: Published article, section 5.2.2, Figure 5 and Table 4, pp. 618 to 621
  • Textual error in the assignment of neuroticism: Published article, p. 621 says Neuroticism loads predominantly on Component 3; Table 4, p. 620, places its strongest loadings on Component 2
  • Test with prompts treated as ground truth and qualitative results: Published article, section 5.3 and Figure 6, pp. 622 to 623
  • Scope as linguistic proxy, declared limitations, and conclusion: Published article, sections 6 to 7, pp. 624 to 627
  • Questions, prompts, scales, and final 40-item version: Published article, Appendices A to F, pp. 627 to 635
  • Year, volume, pages, DOI, and license of the definitive version: ACL Anthology 2025.cl-2.6 and journal title page, verified 15 Jul 2026
  • Absence of identified reproducible artifact: Published article contains model links but no code/data repository or data availability statement; author, ACL, arXiv and GitHub searches on 15 Jul 2026 found no official artifact