Predicting the Big Five Personality Traits in Chinese Counselling Dialogues Using Large Language Models

Evaluation and psychometric validity2024arXivApproved editorial review

Authors: Yang Yan, Lizhi Ma, Anqi Li, Jingsong Ma, Zhenzhong Lan

Keywords: Computation and Language, Artificial Intelligence, Computational Psychometrics, Personality Trait Prediction

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

5
Authors
8
Findings
17
Limitations
21
Evidence

Editorial summary

English

Yan and colleagues ask whether an LLM can estimate Big Five traits from Chinese counselling dialogues. Their procedure does not directly predict five numbers: it places each session into a client, counsellor, or observer role, administers all 60 BFI-2 items to the model, extracts Likert choices with a regex, and computes OCEAN using the inventory scoring rules. The criterion is the BFI-2 completed by each client before their first session. The corpus contains 853 sessions and 65,347 utterances from 82 adults and nine counsellors, although the abstract and introduction say 83 clients. The authors allocate 611 sessions to training and 242 to validation but do not say that the split is client-disjoint, even though each person contributes roughly ten sessions sharing one trait label. In zero-shot evaluation, combining the client role with the questionnaire greatly improves over direct prediction: Qwen1.5-110B rises from mean PCC 0.172 to 0.426 and DeepSeek-Chat from 0.080 to 0.395. The questionnaire without a role already accounts for most of the gain (0.319 and 0.284); role-play alone does not help. Across 21 models, Qwen1.5-110B and Gemini-1.5-Pro reach mean 0.425, while several smaller models are near zero. Roles closer to the dialogue perform better on average: client exceeds counsellor, observer, and no-role, though not for every trait. With 30% of the dialogue, correlations are already significant for the tuned model (mean PCC 0.510), but they continue to rise to 0.631 at 90%; Qwen obtains 0.331 at 30% versus 0.425 with the full context. Thus, 30% is a significance threshold, not evidence of equivalent or practically sufficient performance. To align Llama-3-8B, the authors choose generated responses with the lowest error as preferred and the highest-error responses as rejected, then apply DPO with an SFT constraint. Llama-3-8b-BFI reaches PCC 0.692, 0.554, 0.569, 0.448, and 0.648 for O, C, E, A, and N, averaging 0.582. The reported relative gains of 130.95% over its base (0.252) and 36.94% over Qwen (0.425) are arithmetically correct, while the absolute gains are 0.330 and 0.157. The full matrix contains large cross-trait correlations, for example, predicted O correlates 0.522 with actual E and -0.455 with actual A, limiting discriminant validity. Error analysis provides plausible examples of extracting states and also of misunderstanding, stereotyping, and safety refusal; calling this “content validity” does not replace expert psychometric assessment. Reliability is also ambiguous: the paper reports one alpha per model, trait-level kappas in one table, and PCC across ten runs in another without sufficiently defining units and calculations. Consent, compensation, ethics approval, and offline-only use are documented. The evidence supports that decomposing the task into BFI responses and tuning on error-labelled preferences yields moderate correlations within this corpus; it does not establish an unbiased or safe automated assessment for clinical use.

Español

Yan y colaboradores preguntan si un LLM puede estimar Big Five a partir de diálogos chinos de orientación psicológica. Su procedimiento no predice directamente cinco números: inserta cada sesión en un rol de cliente, consejero u observador, administra al modelo los 60 ítems BFI-2, extrae por regex las opciones Likert y calcula OCEAN con el scoring del inventario. El criterio es el BFI-2 que cada cliente completó antes de su primera sesión. El corpus contiene 853 sesiones y 65.347 intervenciones de 82 adultos y nueve consejeros, aunque resumen e introducción dicen 83 clientes. Se asignan 611 sesiones a entrenamiento y 242 a validación; el artículo no indica que la separación sea por cliente, pese a que cada persona aporta en promedio unas diez sesiones y comparte la misma etiqueta de rasgo. En zero-shot, combinar rol de cliente y cuestionario mejora mucho frente a pedir una predicción directa: Qwen1.5-110B pasa de PCC medio 0,172 a 0,426 y DeepSeek-Chat de 0,080 a 0,395. El cuestionario sin rol ya explica la mayor parte de la mejora (0,319 y 0,284); el rol solo no ayuda. Entre 21 modelos, Qwen1.5-110B y Gemini-1.5-Pro alcanzan 0,425 de media, mientras varios modelos pequeños quedan cerca de cero. Los roles próximos al diálogo funcionan mejor: cliente supera a consejero, observador y sin rol en promedio, aunque no en cada rasgo. Con 30% del texto, las correlaciones ya son significativas para el modelo afinado (PCC medio 0,510), pero continúan creciendo hasta 0,631 al 90%; para Qwen, 30% obtiene 0,331 frente a 0,425 con todo el contexto. Por ello, 30% es un umbral de significación, no evidencia de equivalencia o suficiencia práctica. Para alinear Llama-3-8B, los autores eligen como preferidas las respuestas generadas con menor error y como rechazadas las de mayor error, y aplican DPO con una restricción SFT. Llama-3-8b-BFI alcanza PCC 0,692, 0,554, 0,569, 0,448 y 0,648 para O, C, E, A y N, media 0,582. El aumento relativo de 130,95% frente a su base (0,252) y 36,94% frente a Qwen (0,425) es aritméticamente correcto, aunque la mejora absoluta es 0,330 y 0,157. La matriz completa muestra correlaciones cruzadas grandes, por ejemplo, la O predicha correlaciona 0,522 con E real y -0,455 con A real, lo que limita la validez discriminante. El análisis de errores aporta ejemplos plausibles de extracción de estados y también de malinterpretación, estereotipado y rechazos de seguridad; llamarlo “validez de contenido” no sustituye una evaluación psicométrica por expertos. La fiabilidad también queda ambigua: se informa alfa por modelo, kappas por rasgo en una tabla y PCC de diez ejecuciones en otra, sin describir suficientemente unidades y cálculo. El consentimiento, remuneración, aprobación ética y uso offline están documentados. La evidencia respalda que convertir el problema en respuestas BFI y afinar con etiquetas de error produce correlaciones moderadas dentro de este corpus; no prueba una evaluación automática imparcial ni segura para uso clínico.

Research question

Can LLMs estimate OCEAN from real psychological counseling sessions in Chinese, what role do role-play, BFI decomposition, amount of context, and model capacity play, and how much does a DPO plus SFT alignment improve?

Method

Each session is transformed into a chat history and the LLM adopts the role of client, counselor, observer, or none. It answers with option and reason the 60 BFI-2 items; a regex extracts the option and the scoring produces five traits. Predictions are compared with the client's prior BFI-2 using PCC and MAE, roles, 10–100% of context, 21 models, alpha and repetition are studied, and Llama-3-8B is fine-tuned through preferences chosen by minimum/maximum error, DPO, and an SFT loss.

Sample: 853 sessions from 82 adults, 55 women, 19–54 years, M=27.62, SD=5.94, and nine counselors, seven women, 25–45 years, M=34.67, SD=7.45. The text also states 83 clients in other passages. Each client completed BFI-2 once before the first session. 611 sessions are used for training and 242 for validation, with no separation by person indicated. Participants consented, received 300 RMB, and the protocol was approved with reference 20220519LZZ001.

Findings

  • The combination of questionnaire and client role raises the mean PCC of Qwen1.5-110B from 0.172 to 0.426 and of DeepSeek-Chat from 0.080 to 0.395; role alone contributes little or worsens.
  • Llama-3-8b-BFI obtains a mean PCC of 0.582 compared to 0.252 for the base model and 0.425 for Qwen1.5-110B.
  • The client role is usually better than counselor, observer, and no role, although the differences are not uniform across dimensions.
  • At 30% of context the fine-tuned model reaches 0.510, but the improvement continues up to 0.631 at 90%; partial context is not equivalent to full context.
  • More capable models usually perform better, but the pattern by size is not uniform and only some large members of Qwen show significance.
  • The matrix of the fine-tuned model retains moderate-high diagonal correlations and substantial cross-correlations, a sign of imperfect trait separation.
  • DPO with SFT constraint improves the mean from 0.563 to 0.582 compared to DPO without SFT; the gain is 0.019.
  • Anonymization reduces the average in several conditions, although some dimensions rise, so it does not produce a uniform 6% drop.

Limitations

  • The sampling unit is the session, but the BFI criterion is unique per client; treating up to about ten sessions per person as independent observations may inflate significance and precision.
  • It is not stated that training and validation are separated by client. If a person appears in both, there is leakage of identity, style, and label, especially relevant for the fine-tuned model.
  • Client counts are inconsistent: 83 in the summary/introduction and 82 in the methods; some session averages in Table 5 also do not match perfectly.
  • Only the validation set of 242 sessions is described as anonymized; this introduces a distribution shift and leaves ambiguous the treatment of identifiable data used to train and release the model.
  • Correlations are presented with asterisks without a multilevel model, intervals, or correction for the numerous tests; intra-client dependence violates the assumption of independent observations.
  • A correlation with a single self-report does not demonstrate clinical accuracy or absence of bias; the BFI-2 is a criterion with its own measurement error, context, and desirability.
  • Administering the same BFI-2 to the LLM turns the task into imitation of questionnaire responses and may exploit stereotypes or item associations, not general psychological inference.
  • The qualitative analysis of hits and errors is labeled as content validity without expert judges, rubric, systematic sampling, or inter-rater agreement.
  • Reliability is not sufficiently defined: alpha appears aggregated by model, Table 4 reports kappa, and Table 6 reports PCC of ten runs, without reconciling the measures or giving individual stability.
  • The claim that 30% of context is sufficient confuses significance with equivalence; results improve materially when using 80–100% and there is no non-inferiority margin.
  • The MAE<1 threshold is called significant without statistical justification, and boxplots do not replace a table of errors by trait, person, and session.
  • The matrices show strong negative off-diagonal correlations that question discriminant validity, but the article emphasizes only the diagonal.
  • The gains of 130.95% and 36.94% are relative percentages over PCC means; their formulation magnifies absolute changes of 0.330 and 0.157.
  • DPO pairs are chosen using error against the label, but it is not described how many there are, their distribution, possible duplicates per client, or independent evaluation of the generated reasoning.
  • The corpus comes from a single platform, language, and region, with relatively young adults; there is no external, longitudinal, cross-cultural, or real clinical decision validation.
  • Comparing 21 APIs and checkpoints without dates, sampling parameters, number of retries, complete failure handling, and cost limits the reproducibility of the ranking.
  • Although the code and the model are linked, the sensitive dialogues are not published, so third parties cannot fully audit splits, labels, leakage, anonymization, or results.

What the study does not establish

  • It does not demonstrate that LLMs measure a person's true personality or that they surpass a professional psychometric evaluation.
  • It does not demonstrate that the system is unbiased; the 'unbiased' method conclusion is not supported by analysis by gender, age, client, counselor, or subgroup.
  • It does not demonstrate generalization to new clients if the splits are not explicitly disjoint by person.
  • It does not demonstrate that 30% of a session retains performance equivalent to the full text.
  • It does not demonstrate full discriminant validity, since strong correlations exist with non-target traits.
  • It does not validate clinical deployment, diagnosis, treatment selection, patient monitoring, or replacement of consent or professional judgment.

Traceability

Scope: Full text

Version: arXiv:2406.17287v1 (25 June 2024)

Consulted source: https://arxiv.org/pdf/2406.17287

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • Llama-3-8b-BFI
  • Meta-Llama-3-8B-Instruct
  • Meta-Llama-3-70B-Instruct
  • Qwen1.5-110B-Chat
  • Qwen-72B-Chat
  • Qwen1.5-7B-Chat
  • Qwen2-7B-Instruct
  • GPT-4-turbo
  • DeepSeek-Chat
  • deepseek-llm-67b-chat
  • Gemini 1.5 Pro and Flash
  • Gemini 1.0 Ultra and Pro
  • Qwen Long and Turbo
  • ERNIE Speed and Lite
  • Yi-34B-Chat
  • AquilaChat2-34B
  • InternLM2-Chat-20B
  • Baichuan2-13B-Chat
  • GLM-4-9B-Chat
  • Gemma-1.1-7B-it
  • ChatGLM3-6B-128K

Instruments and metrics

  • Chinese Big Five Inventory-2, 60 items
  • Pearson correlation coefficient
  • Mean absolute error
  • Cronbach alpha
  • Kappa reported for test-retest reliability
  • Ten-run PCC stability analysis
  • Interquartile-range outlier rule
  • Direct Preference Optimization with supervised fine-tuning constraint

Data used

  • 853 Chinese counselling sessions from the authors' platform
  • 65,347 counsellor and client utterances
  • Pre-session BFI-2 profiles from 82 reported adult clients
  • 611-session training split
  • 242-session manually anonymized validation split
  • Model-generated BFI rationales selected by minimum and maximum prediction error

Evidence and location

  • Question, framework, session count, and main claims: arXiv v1, pp. 1–2, Abstract and Introduction
  • Role-play and questionnaire ablation: arXiv v1, p. 3, Table 1
  • Prompt, roles, 60 items, and regex extraction: arXiv v1, pp. 3–5, sections 3.1–3.2
  • PCC, MAE, and unconventional definition of content validity: arXiv v1, p. 5, section 3.3
  • 82 clients, nine counselors, and split by sessions: arXiv v1, p. 5, section 4.1
  • Results by role: arXiv v1, pp. 4 and 6, Table 2 and section 4.2
  • Comparison of 21 models: arXiv v1, pp. 5–6, Table 3 and section 4.2
  • Narrative threshold of 30% context: arXiv v1, p. 6, Figure 2 and section 4.2
  • Errors, rejections, and discrepancies with self-report: arXiv v1, pp. 6–8, section 4.3 and Figure 4
  • Preference construction, DPO, and SFT: arXiv v1, p. 8, section 4.4 and Figure 5
  • Relative gains and declared efficiency: arXiv v1, p. 8, Model Proximity subsection
  • Consent, remuneration, ethics, and offline use: arXiv v1, p. 9, Ethical Considerations
  • Acknowledged limitations, privacy, and absence of benchmark: arXiv v1, p. 9, Limitations
  • Full BFI-2 and scoring: arXiv v1, p. 14, Appendix A.1.1
  • Reliability alpha, kappa, and ten runs: arXiv v1, pp. 15–16, Table 4, Appendix A.4 and Table 6
  • Dialogue statistics and anonymization of 242 sessions: arXiv v1, pp. 15–16, Appendix A.2 and Table 5
  • Complete results by 10–100% context: arXiv v1, p. 17, Table 8
  • DPO+SFT effect and roles with names: arXiv v1, p. 18, Tables 9–10
  • Inverted splits and fine-tuned models ablation: arXiv v1, p. 19, Table 11
  • Diagonal and cross-correlations: arXiv v1, pp. 19–20, Figures 6–11
  • Training hyperparameters: arXiv v1, p. 21, Table 12