Personality prediction from task-oriented and open-domain human–machine dialogues

Personas, identity, and agents2024Scientific ReportsApproved editorial review

Original title: Personality prediction from task-oriented and open-domain human-machine dialogues

Authors: Ao Guo, Ryu Hirai, Atsumoto Ohashi, Yuya Chiba, Yuiko Tsunomori, Ryuichiro Higashinaka

Keywords: Personality Prediction, Human-Machine Dialogue, Big Five, BFI-44, KISS-18, Adult Temperament Questionnaire, BERT, Balanced Accuracy, Domain Shift, Psychometric Validity, Reproducibility

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

6
Authors
12
Findings
21
Limitations
9
Evidence

Editorial summary

English

This work attempts to infer human traits, not machine personality. Amazon Mechanical Turk participants completed BFI-44, IOS, KISS-18 and ATQ measures and held three conversations with one of three systems: a pipeline task system, task-oriented SimpleTOD or open-domain BlenderBot. After filtering, 179, 199 and 186 participants remained. BERT-base-uncased receives only user messages and learns 25 separate targets.

“Personality prediction” here means classifying whether a score is above or below the training-plus-validation median. It does not estimate a continuous score or normative level. The published rule also leaves exact median ties undefined, important because questionnaire totals are discrete. Thresholds change across splits, so .64 balanced accuracy describes a relative, moving label rather than a stable individual assessment.

The strongest results occur in open-domain dialogue: .71 Conscientiousness, .68 Agreeableness, .60 Neuroticism, .61 IOS and .60-.68 for five temperament/social facets. Extraversion reaches .64 in both open-domain and pipeline task dialogue. Many of the 25 outcomes remain near the .50 baseline. The only comparator is a majority classifier whose balanced accuracy is .50 by construction; length, vocabulary, TF-IDF and linear baselines are absent, so BERT's added value is unknown.

The open-domain advantage is not causal. Each condition uses different people, with no reported random assignment or cohort demographics. Domain, architecture, difficulty, turn limit, content, system persona, quality and success rate change together. Open-domain dialogue also contains more user vocabulary (90.2 terms per dialogue versus 70.8 and 58.1) and higher lexical diversity. The comparison is between complete packages and populations.

Cross-domain performance is often near chance, useful evidence of distribution shift. But only this BERT and three historical corpora are tested without domain adaptation; the study cannot show that open-domain data can never aid task prediction. Nor does it establish that one round is sufficient or task success irrelevant: these are absence-of-significance claims without equivalence tests or power. The proposed SimpleTOD free-expression mechanism rests on success rates and scatterplots, not causal analysis.

Statistical inference is not reproducible: the article names a “non-parametric t-test”, a contradictory label, and compares overlapping cross-validation estimates without identifying the test statistic. Questionnaire scoring, reverse coding, reliability and distributions are also omitted. Data and code are available only “upon reasonable request”; no public release was found through Nature, PMC, the author page or author GitHub.

The faithful conclusion is modest: in these AMT samples and historical systems, some relative trait groups leave textual signals recoverable by BERT, especially in unconstrained dialogue. Continuous validity, operational utility, fairness, personalization benefit and generalization to current LLMs are not validated. Balanced accuracy of .60-.71 still implies substantial error, and the study never deploys adaptive decisions to test benefit or harm.

Español

Este trabajo intenta inferir rasgos humanos, no personalidad de la máquina. Personas reclutadas en Amazon Mechanical Turk completaron BFI-44, IOS, KISS-18 y ATQ y mantuvieron tres conversaciones con uno de tres sistemas: un pipeline orientado a tareas, SimpleTOD orientado a tareas o BlenderBot abierto. Tras filtros quedaron 179, 199 y 186 participantes, respectivamente. BERT-base-uncased recibe solo los mensajes del usuario y aprende por separado 25 objetivos.

“Predecir personalidad” significa aquí clasificar si una puntuación está por encima o por debajo de la mediana de training+validation. No se predice el valor continuo ni un nivel normativo. Además, la regla publicada deja sin definir los empates exactamente en la mediana, relevantes porque los cuestionarios producen puntuaciones discretas. El umbral cambia entre particiones, por lo que una balanced accuracy de 0,64 describe una etiqueta relativa y móvil, no una evaluación individual estable.

Los mejores resultados aparecen en diálogo abierto: 0,71 para Conscientiousness, 0,68 Agreeableness, 0,60 Neuroticism, 0,61 IOS y 0,60-0,68 para cinco facetas de temperamento/sociales. Extraversion llega a 0,64 tanto en diálogo abierto como en el pipeline de tareas. Muchas de las 25 variables se mantienen cerca del baseline de 0,50. El único comparador es un clasificador mayoritario cuya balanced accuracy es 0,50 por construcción; faltan baselines de longitud, vocabulario, TF-IDF o modelos lineales que indiquen cuánto añade BERT.

No puede atribuirse causalmente la ventaja al diálogo abierto. Cada condición usa personas distintas y no se informa asignación aleatoria ni demografía por cohorte. Cambian simultáneamente dominio, arquitectura, dificultad, límite de turnos, contenido, persona del sistema, calidad y tasa de éxito. El diálogo abierto también contiene más vocabulario del usuario (90,2 términos por diálogo frente a 70,8 y 58,1) y mayor diversidad léxica. La comparación mide paquetes completos y distribuciones diferentes.

La transferencia entre dominios suele acercarse al azar, un hallazgo útil sobre distribution shift. Pero solo se prueba este BERT y estos tres corpus históricos, sin adaptación de dominio; no demuestra que datos abiertos nunca puedan ayudar en tareas. Tampoco se demuestra que una ronda sea “suficiente” o que el éxito de la tarea sea irrelevante: esas conclusiones se basan en pocas diferencias significativas, sin equivalence tests o potencia. La explicación de que SimpleTOD limita la libre expresión se apoya en tasa de éxito y scatterplots, no en una prueba causal.

La inferencia estadística no es reproducible: el artículo dice “non-parametric t-test”, una denominación contradictoria, y compara estimaciones de cross-validation solapadas. No se identifica el test ni su estadístico. También faltan scoring, reverse coding, fiabilidad y distribuciones de los cuestionarios. Los datos y el código se ofrecen solo “upon reasonable request”; no se localizó release público en Nature, PMC, la página del autor o su GitHub.

La conclusión fiel es modesta: en estas muestras AMT y sistemas de 2023, algunos grupos relativos de rasgos dejan señales textuales recuperables por BERT, sobre todo cuando la conversación es abierta. No se validan puntuaciones continuas, utilidad operativa, fairness, beneficio de personalización ni generalización a LLM actuales. Una tasa de 0,60-0,71 implica errores sustanciales y el estudio no despliega decisiones adaptativas para comprobar si ayudan o dañan.

Research question

With what balanced accuracy can BERT separate scores of 25 human traits above or below the median using messages in task-oriented or open human-machine dialogues, and how much do models transfer between those domains?

Method

AMT study with four questionnaires and three rounds per participant in a single condition: ConvLab-2 pipeline, SimpleTOD E2E or BlenderBot 400M open. After filters, 179/199/186 people remain. Each trait is dichotomized by the median of training+validation. BERT-base-uncased classifies user text in 10-fold cross-validation repeated ten times; balanced accuracy and a majority baseline are used. Ten traits with at least 0.60 enter the cross-domain 8:1:1 comparison averaged over 100 seeds. Longformer, rounds and success/failure are analyzed additionally.

Sample: 610 AMT participants initially: 204 pipeline, 205 E2E and 201 open-domain. After excluding for completing three rounds in less than five minutes or having any round with fewer than ten unique words, 179, 199 and 186 remain (564 total). Regions limited to English-speaking countries, more than 100 HIT and approval >95%, payment USD10. Age, gender, country, education and demographic balance across conditions are not reported.

Findings

  • Open dialogue obtains balanced accuracy 0.71 Conscientiousness, 0.68 Agreeableness, 0.60 Neuroticism and 0.61 IOS.
  • In open also reported 0.64 Frustration, 0.60 Activation Control, 0.65 Attentional Control, 0.62 Sociability and 0.68 Neutral Perceptual Sensitivity.
  • Extraversion reaches 0.64 in both open and task-oriented pipeline.
  • Most other traits remain near 0.50-0.60.
  • Open-domain has user vocabulary 90.2 and diversity 0.42 versus 70.8/0.30 pipeline and 58.1/0.35 E2E.
  • The success rate is 47% pipeline and 28% E2E.
  • Transfer O→P remains near chance: 0.52 C, 0.52 E, 0.51 A, 0.50 N and 0.56 IOS.
  • Transfer P→O gives 0.57 C, 0.62 E, 0.54 A, 0.49 N and 0.57 IOS: weak, but not always equal to chance.
  • Adding system text usually reduces Longformer performance.
  • Few differences are detected between rounds and between successful/failed dialogues, without demonstrating equivalence.
  • The article acknowledges that 0.60 is not high and that errors may reduce satisfaction.
  • Code and data are request-only; no public release was located.

Limitations

  • Continuous traits are reduced to relative high/low categories and magnitude is lost.
  • The threshold changes per split and does not define what happens with values equal to the median.
  • Each condition uses distinct cohorts without random assignment or reported demographics.
  • Domain, architecture, task, difficulty, length, persona, quality and success change at the same time.
  • Open dialogue offers more text and diversity, a direct confounder of prediction.
  • The only baseline is majority with balanced accuracy 0.50 by construction.
  • There are no lexical, linear, TF-IDF or length/vocabulary baselines.
  • The filter removes participants for low vocabulary, a signal that the predictor then uses.
  • Scoring, reverse coding, reliability, missingness or distributions of the questionnaires are not reported.
  • IOS and social skills are not equivalent to broad personality traits.
  • The expression 'non-parametric t-test' does not identify a valid statistical procedure.
  • Cross-validation repetitions share data and are not independent observations.
  • Generalization intervals per participant or calibration are not shown.
  • The cross-domain comparison selects post hoc ten traits with at least 0.60.
  • No domain adaptation or external corpus is tested.
  • Absence of significant differences does not demonstrate sufficiency or irrelevance.
  • The causal explanation of SimpleTOD is based on descriptives.
  • There is no evaluation of fairness, privacy, consent for profiling or error by subgroup.
  • Utility for adaptation is not tested in a deployed system.
  • Code, data, splits, seeds, versions and checkpoints are not public.
  • Dialogue systems are historical and do not represent current LLMs.

What the study does not establish

  • It does not predict continuous scores or clinical/normative levels of personality.
  • It does not demonstrate that open dialogue causes better inference.
  • It does not demonstrate that 0.60-0.71 is sufficient for personalization.
  • It does not test benefit in satisfaction, trust or task success.
  • It does not demonstrate equivalence between one and three rounds.
  • It does not demonstrate that task success is irrelevant.
  • It does not demonstrate that SimpleTOD causes lower free expression.
  • It does not demonstrate universal impossibility of transfer between domains.
  • It does not demonstrate that BERT outperforms simple lexical baselines.
  • It does not demonstrate reliability and equivalence of labels across cohorts.
  • It does not demonstrate fairness across groups.
  • It does not allow reproducing figures with public artifacts.
  • It does not automatically generalize to current LLMs, voice or non-AMT populations.
  • It does not study or demonstrate AI personality.

Traceability

Scope: Full text

Version: Scientific Reports 14:3868 version of record, published online 16 February 2024

Consulted source: https://www.nature.com/articles/s41598-024-53989-y

Review: Codex full-text, visual, psychometric, statistical and artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • BERT-base-uncased personality classifiers
  • Longformer-base-4096 user/system-text analysis
  • ConvLab-2 pipeline task-oriented dialogue system
  • GPT-2-based SimpleTOD end-to-end task-oriented dialogue system
  • Distilled BlenderBot 400M open-domain dialogue system

Instruments and metrics

  • Big Five Inventory 44 (BFI-44)
  • Inclusion of Other in the Self (IOS) Scale
  • Kikuchi's Scale of Social Skills (KISS-18)
  • Adult Temperament Questionnaire (ATQ)
  • Above/below training-validation median labels
  • Balanced accuracy
  • Repeated participant-level 10-fold cross-validation

Data used

  • 537 retained pipeline task dialogues from 179 participants (request-only)
  • 597 retained SimpleTOD task dialogues from 199 participants (request-only)
  • 558 retained BlenderBot open-domain dialogues from 186 participants (request-only)
  • MultiWOZ 2.1 task goals
  • Four-questionnaire participant trait labels (request-only)

Evidence and location

  • Constructs, dichotomization, model, metric and baseline: Version of record pp. 3-4, Approach, Personality labeling, BERT-based model and Evaluation metric
  • Recruitment, systems, filters and statistics: Version of record pp. 4-6, Data collection and Tables 1-2
  • Configuration, repeated CV and main results: Version of record pp. 6-7, Experiments and Figure 2
  • Transfer between domains: Version of record pp. 7-8, Table 3 and cross-comparison section
  • System text, rounds, success and E2E mechanism: Version of record pp. 8-11, Figures 3-6 and Analyses
  • Cautions and operational recommendation: Version of record pp. 10-11, Conclusion and future work
  • Request-only availability: Version of record p. 11, Data availability
  • Absence of public release: Nature, PMC, author site and GitHub guoao8644 title/DOI searches audited 16 Jul 2026
  • Complete report: reports/verification/article-212-dialogue-personality-prediction-validity-and-artifact-audit.json