PSYDIAL: Personality-based Synthetic Dialogue Generation Using Large Language Models

Personas, identity, and agents2024ACL AnthologyApproved editorial review

Authors: Ji-Eun Han, Jun-Seok Koh, Hyeon-Tae Seo, Du-Seong Chang, Kyung-Ah Sohn

Keywords: synthetic dialogue generation, personality-based dialogue, Big Five extraversion, multilingual NLP, Korean language, conversational AI, data augmentation

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

5
Authors
26
Findings
40
Limitations
18
Evidence

Editorial summary

English

PSYDIAL proposes a five-stage pipeline for generating personality-conditioned synthetic Korean dialogues: trait assignment, PersonaChat profile selection, GPT-4 generation, automatic filtering, and up to three regeneration rounds. Although the paper frames the task around Big Five personality, the experiment uses only one dimension, Extraversion, reduced to extrovert and introvert poles for each of two speakers. Of 4,000 initial dialogues, the filter rejects 1,051 for profile, 208 for personality, and one for style; regeneration yields 4,305 cumulative attempts and Table 4 reports 2,928 positives. Table 1, however, sums to 2,932 final dialogues, 715 E/E, 685 E/I, 763 I/E, and 769 I/I, an unexplained four-item discrepancy. Dialogues average 8.16 turns and 33.25 syllable-level tokens. Selected profiles expose stereotypical associations: travel, dancing, football, hiking, and swimming for extroversion, versus reading, video games, painting, being alone, and also hiking for introversion. A t-SNE plot suggests post-filter separation, but no separation metric or statistical test is reported. For downstream evaluation, the authors compare KoGPT2, KoBART, Kolang-T5, and KoDialoGPT-v0 under five settings: pretrained, personality-prompted, fine-tuned for three epochs on HuLiC, fine-tuned for three epochs on PSYDIAL, and PSYDIAL fine-tuning plus a personality prompt. PSYDIAL-trained models with prompting reach P-ACC of 0.881, 0.864, and 0.864 for KoGPT2, KoBART, and Kolang-T5, versus 0.653, 0.664, and 0.625 without the prompt; most BLEU and ROUGE scores also improve. This supports the narrower claim that training on labeled dialogues and restating the label in the prompt helps reproduce synthetic classes within this protocol. It does not establish broad personality, human realism, or psychometric validity. P-ACC comes from KLUE RoBERTa-base fine-tuned for five epochs on PSYDIAL, yet no train/test split or safeguard against leakage across generation, classification, and evaluation is described. There is also no human evaluation of Korean naturalness or personality; the authors themselves note English-like direct-translation expressions. The code audit further exposes circularity: the history used by GPT-4 to classify a dialogue still contains the explicit trait prompt, while the profile filter receives the selected profile rather than the dialogue. The official repository publishes three scripts and two profile CSV files, but not the final PSYDIAL corpus, fine-tuned models, splits, result artifacts, a dependency manifest, or a license. The paper offers a useful initial recipe for synthetic conversational data in a lower-resource setting, but the published evidence does not independently verify data quality, reproduce the full evaluation, or support generalization beyond binary Extraversion in Korean.

Español

PSYDIAL propone un pipeline de cinco etapas para generar diálogos sintéticos en coreano condicionados por personalidad: asignación del rasgo, selección de un perfil de PersonaChat, generación con GPT-4, filtrado automático y hasta tres regeneraciones. Aunque el artículo habla de personalidad basada en Big Five, el experimento solo usa una dimensión, extraversión, reducida a dos polos, extrovertido e introvertido, para cada uno de dos interlocutores. De 4.000 diálogos iniciales, el filtro rechaza 1.051 por perfil, 208 por personalidad y uno por estilo; las regeneraciones acumulan 4.305 intentos y la Tabla 4 declara 2.928 positivos. Sin embargo, la Tabla 1 suma 2.932 diálogos finales, 715 E/E, 685 E/I, 763 I/E y 769 I/I, una discrepancia interna de cuatro casos sin explicar. Los diálogos tienen una media de 8,16 turnos y 33,25 tokens a nivel silábico. Los perfiles escogidos revelan asociaciones estereotípicas: viaje, baile, fútbol, senderismo y natación aparecen para extroversión, mientras lectura, videojuegos, pintura, soledad y también senderismo aparecen para introversión. La visualización t-SNE sugiere separación tras el filtrado, pero no se aporta una métrica de separación ni una prueba estadística. Para evaluar utilidad, los autores comparan KoGPT2, KoBART, Kolang-T5 y KoDialoGPT-v0 en cinco configuraciones: preentrenado, con prompt de personalidad, ajustado tres épocas con HuLiC, ajustado tres épocas con PSYDIAL y este último más prompt. Los modelos ajustados con PSYDIAL y prompt alcanzan P-ACC de 0,881, 0,864 y 0,864 para KoGPT2, KoBART y Kolang-T5, frente a 0,653, 0,664 y 0,625 sin el prompt; también mejoran la mayoría de BLEU y ROUGE. Esto respalda que entrenar con diálogos etiquetados y repetir la etiqueta en el prompt facilita reproducir las clases sintéticas dentro de este protocolo. No demuestra personalidad amplia, realismo humano ni validez psicométrica. P-ACC procede de un KLUE RoBERTa-base ajustado durante cinco épocas sobre PSYDIAL, pero no se describen particiones de entrenamiento y prueba ni protección frente a fuga entre generador, clasificador y evaluación. Tampoco hay evaluación humana de naturalidad o personalidad coreana; los propios autores reconocen expresiones parecidas a traducciones directas del inglés. La auditoría del código agrava la circularidad: el mismo historial contiene el prompt con los rasgos explícitos cuando GPT-4 vuelve a clasificar el diálogo, y el filtro de perfil recibe el perfil seleccionado, no el diálogo. El repositorio oficial publica tres scripts y dos CSV de perfiles, pero no el corpus PSYDIAL final, los modelos ajustados, las particiones, los resultados, un manifiesto de dependencias ni una licencia. El trabajo aporta una receta inicial útil para crear datos conversacionales sintéticos en un recurso lingüístico reducido, pero la evidencia publicada no permite verificar de forma independiente su calidad, reproducir la evaluación completa ni generalizar sus resultados más allá de extraversión binaria en coreano.

Research question

Can a pipeline based on GPT-4 prompting generate and filter Korean synthetic dialogues that express poles of extraversion, and does fine-tuning conversational models with those data improve automatic generation evaluated by overlap, perplexity, and a personality classifier?

Method

The pipeline assigns Person1 and Person2 one of two poles of extraversion, selects via GPT-4 a PersonaChat profile sentence compatible with Person1, generates an informal dialogue in Korean, and asks the same GPT-4 to classify profile, personality, and style. Negative cases are regenerated up to three times. Four combinations are constructed: E/E, E/I, I/E, and I/I. For profile analysis, frequent terms are counted and embeddings from a Korean sentence-transformer are visualized with t-SNE. Utility is studied with KoGPT2, KoBART, and Kolang-T5 fine-tuned for three epochs, plus KoDialoGPT-v0 as a reference, in five configurations that cross pretraining, HuLiC, PSYDIAL, and personality prompt. BLEU-2, ROUGE-1/2/L, 3-gram perplexity estimated with KoGPT2, and P-ACC of a KLUE RoBERTa-base fine-tuned for five epochs with PSYDIAL are reported.

Sample: The pipeline starts from 4,000 initial generations, 1,000 per each combination E/E, E/I, I/E, and I/I. After filtering and up to three regenerations, Table 1 reports 715, 685, 763, and 769 dialogues, respectively, which sum to 2,932; Table 4 reports 2,928 positives. The declared average length is 8.16 turns and 33.25 syllabic tokens. No humans participate and no human dialogues are annotated for personality. HuLiC is used as a comparative fine-tuning corpus, but the article does not document reproducible sizes or partitions for PSYDIAL in this evaluation.

Findings

  • The study operationalizes personality solely as binary extraversion for two speakers; it does not generate or evaluate the other four Big Five dimensions.
  • The four speaker combinations are E/E, E/I, I/E, and I/I, with 1,000 initial attempts per combination.
  • Table 1 declares 715, 685, 763, and 769 final dialogues; these figures sum to 2,932.
  • Table 4 declares 2,928 accumulated positives, four fewer than Table 1, with no textual reconciliation.
  • Of the 4,000 initial dialogues, 1,051 fail the profile filter, 208 the personality filter, and one the style filter; 2,740 pass directly.
  • Three regeneration rounds add 138, 37, and 13 positives, with 4,305 accumulated attempts.
  • The dialogues have a mean of 8.16 turns and 33.25 tokens measured at the syllabic level.
  • The selected extroverted profiles are associated with travel, dancing, football, hiking, and swimming; the introverted ones with reading, video games, painting, being alone, and also hiking.
  • The presence of hiking in both poles shows that profile selection does not produce a simple or exclusive semantic separation.
  • Figure 2 shows a greater visual separation after filtering, but does not quantify cluster quality, error, or significance.
  • The prompts are designed in English to generate conversations in Korean, and the article recognizes expressions similar to direct translations from English.
  • The personality prompt explicitly adds each interlocutor's label to the same context that generates the dialogue.
  • In the official code, that labeled context remains in the history when GPT-4 subsequently classifies personality, so the filtering is not blind to the treatment.
  • The separate script's profile filter receives only the selected profile and the expected label, not the dialogue whose coherence it should verify.
  • In the pretrained configuration, P-ACC is 0.481 for KoGPT2, 0.463 for KoBART, and 0.565 for Kolang-T5.
  • Fine-tuning with HuLiC does not produce a uniform improvement in personality: P-ACC remains at 0.505, 0.531, and 0.518.
  • Fine-tuning with PSYDIAL raises P-ACC to 0.653, 0.664, and 0.625 for KoGPT2, KoBART, and Kolang-T5.
  • Adding the personality prompt to the models fine-tuned with PSYDIAL raises P-ACC to 0.881, 0.864, and 0.864.
  • The article's formulation about an increase of up to 88% corresponds to reaching 88.1% accuracy, not to an improvement of 88 percentage points.
  • Most BLEU and ROUGE improve when moving from PSYDIAL without prompt to PSYDIAL with prompt, consistent with an additional explicit signal.
  • Kolang-T5 perplexity worsens from 15.223 to 16.521 when adding the prompt, so not all metrics improve.
  • P-ACC is computed with a KLUE RoBERTa-base fine-tuned for five epochs on PSYDIAL, the same synthetic domain it evaluates.
  • The paper does not describe a human evaluation of naturalness, coherence, diversity, or perceived personality.
  • The official repository contains the generation, filtering, and regeneration code and two input CSVs, but not the final corpus or evaluation artifacts.
  • The README references main_filtering.py and main_regeneration.py, but the published files are named main_filter.py and main_regenerate.py.
  • The code defines a moderation function, but does not invoke it in the published pipeline.

Limitations

  • Reducing personality to a single binary trait prevents sustaining claims about Big Five or personality in a broad sense.
  • The extroverted and introverted poles are treated as discrete and symmetric classes, without intensity, facets, ambiversion, or uncertainty.
  • The profile is chosen from PersonaChat in English and the dialogue is generated in Korean, introducing cultural and linguistic transfer that is not evaluated.
  • Profile selection reproduces stereotypical associations between activities and extraversion.
  • It is not measured whether the selected profiles are psychologically valid or whether Korean people recognize the traits.
  • The same GPT-4 generates and judges the data, without independent judges or human annotation.
  • The judge keeps in the history the prompt that explicitly reveals the assigned personality, creating a direct path for label leakage.
  • The published profile filter evaluates the profile itself, not its realization in the dialogue.
  • Agreement between filters, judge errors, false positives, or false negatives are not quantified.
  • Regenerations reuse the same model and criteria, so they may intensify stereotypes and judge self-consistency.
  • The t-SNE visualization is exploratory, sensitive to parameters, and does not substitute a quantitative measure of separability.
  • t-SNE parameters, stability across seeds, or validation outside the two-dimensional projection are not documented.
  • Table 1 sums 2,932 cases and Table 4 reports 2,928 positives; the published corpus is not available to resolve the difference.
  • The article does not clearly describe training, validation, and test partitions for the generative models.
  • Nor does it describe the partition used to fine-tune and evaluate the P-ACC classifier on PSYDIAL.
  • Without partitions by profile, dialogue, or template, P-ACC may exploit overlap, generator styles, or label leakage.
  • The evaluation classifier is trained with the same synthetic dataset and binary taxonomy, so the metric is circular with respect to the objective.
  • BLEU and ROUGE reward overlap with synthetic reference responses in an open dialogue task with many valid responses.
  • Perplexity is computed with KoGPT2 3-gram according to the description and is not justified as a measure of human naturalness.
  • There are no confidence intervals, deviations, statistical tests, or replications to support the term significant improvements in the abstract.
  • Fine-tuning is summarized almost solely by the number of epochs; learning rate, batch, optimizer, seeds, checkpoints, and selection criterion are missing.
  • Comparable decoding configurations for the evaluated outputs of the fine-tuned models are not reported.
  • KoDialoGPT-v0 is not fine-tuned under all conditions, which limits its comparability with the other three models.
  • Generators other than GPT-4 or robustness to endpoint versions are not evaluated.
  • The gpt-4 alias in the code does not fix an exact snapshot and does not allow reproducing the original behavior.
  • The code uses the legacy openai.ChatCompletion interface and does not include dependency versions.
  • Random selection of profiles and descriptors does not set a seed.
  • The scripts do not robustly create all output directories and require manual reconstruction of the environment.
  • The repository does not contain requirements.txt, pyproject.toml, lockfile, environment configuration, or complete execution instructions.
  • The README uses file names that do not match two actual scripts.
  • The final PSYDIAL corpus, partitions, model outputs, checkpoints, and metric results are not in the audited repository.
  • There is no visible license for code or data, which leaves academic reuse and redistribution uncertain.
  • Rights and conditions for PersonaChat, HuLiC, GPT-4 outputs, or derived data are not discussed.
  • The moderation function is defined but not used, and no audit of toxicity, privacy, or harmful content is reported.
  • There is no human evaluation of natural Korean, cultural adequacy, coherence across turns, diversity, or conversational utility.
  • The authors acknowledge expressions similar to direct translation from English, but do not quantify their prevalence or impact.
  • The synthetic dialogues are not compared with equivalent human conversations labeled by personality.
  • Memorability, consistency in long conversations, behavior outside the prompt, or interaction with users are not analyzed.
  • The study does not present a substantive ethics section on stereotypes, data consent, risks of personification, or manipulative uses.
  • Generalization to other dimensions, languages, and tasks is raised as potential, not evaluated empirically.

What the study does not establish

  • It does not demonstrate that GPT-4 or the Korean models possess psychological personality.
  • It does not demonstrate control of the five Big Five traits; it only studies binary extraversion.
  • It does not demonstrate that activity profiles are valid or non-stereotypical indicators of extraversion.
  • It does not demonstrate that the dialogues are more human, because there is no direct human comparison.
  • It does not demonstrate naturalness or cultural adequacy for Korean speakers.
  • It does not demonstrate that filtering identifies personality from the dialogue without access to the label.
  • It does not demonstrate that the t-SNE visual separation is statistically robust.
  • It does not demonstrate that 2,932 or 2,928 is the correct size of the final corpus.
  • It does not demonstrate generalization to new profiles, prompts, generators, models, or domains.
  • It does not demonstrate that P-ACC measures personality rather than generator and synthetic label signals.
  • It does not demonstrate a statistically significant improvement by not providing tests, variance, or intervals.
  • It does not demonstrate that adding the prompt improves all metrics; Kolang-T5 worsens in perplexity.
  • It does not demonstrate superiority over a comparable human personality corpus.
  • It does not allow reproducing the construction of the corpus or the evaluation tables from the published artifacts.
  • It does not establish that the pipeline is safe, ethical, or suitable for deployments with users.
  • It does not verify the claim of applicability to non-conversational tasks or other languages.

Traceability

Scope: Full text

Version: LREC-COLING 2024 final paper, pp. 13321-13331, 11 pages; official PSYDIAL repository main audited at commit 08cd9a2c8037d65adf4b5ec4e77b3c072ed2982b; pipeline code and source profile files present, final 2.9K-dialogue corpus absent

Consulted source: https://aclanthology.org/2024.lrec-main.1166.pdf

Review: Codex full-text, bilingual-fidelity, visual, bibliographic, synthetic-data, leakage, Korean-language, automatic-evaluation, code-artifact, reproducibility, release and ethics audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4 mediante el alias no versionado gpt-4 para selección, generación, filtrado y regeneración
  • KoGPT2, 125 millones de parámetros
  • KoBART, 123 millones de parámetros
  • Kolang-T5, 225 millones de parámetros
  • KoDialoGPT-v0, 125 millones de parámetros
  • KLUE RoBERTa-base ajustado como clasificador de P-ACC
  • Sentence-transformer coreano para embeddings de la visualización t-SNE

Instruments and metrics

  • Una dimensión Big Five: extraversión frente a introversión
  • Prompts de perfil, personalidad, personaje y estilo
  • Filtrado GPT-4 de perfil, personalidad y estilo
  • BLEU-2
  • ROUGE-1, ROUGE-2 y ROUGE-L
  • Perplejidad 3-gram con KoGPT2
  • P-ACC con clasificador KLUE RoBERTa-base
  • Embeddings de sentence-transformer y t-SNE
  • Frecuencias de palabras de perfiles seleccionados

Data used

  • PSYDIAL: aproximadamente 2,9 mil diálogos sintéticos coreanos etiquetados por la combinación binaria de extraversión de dos interlocutores
  • PersonaChat como fuente de frases de perfil en inglés
  • HuLiC como corpus humano coreano de chit-chat para la condición de ajuste comparativa
  • 5_personality.csv del repositorio, con descripciones de polos Big Five
  • personaChat_personality_origin.csv del repositorio, con perfiles fuente; el corpus PSYDIAL final no está publicado en la revisión auditada

Evidence and location

  • Bibliographic record and official abstract: ACL Anthology 2024.lrec-main.1166; LREC-COLING 2024, pp. 13321-13331
  • Full audited source: .cache/editorial-sources/article-099/source.pdf; official ACL PDF; 11 pages; sha256 ac43d5a12f565467fc540cdb60263b6c3a78229806d5fc50c37f794eae20e3b7
  • Five-stage pipeline: Full text pp. 13323-13326, sections 3.1-3.4 and Figure 1
  • Profile, personality, character, and style prompts: Full text pp. 13324-13325, sections 3.3.1-3.3.4
  • Corpus distribution and length: Full text pp. 13326-13327, section 4.1 and Tables 1-2
  • Profiles and lexical stereotypes: Full text p. 13327, section 4.2.1 and Table 3
  • Filtering, regenerations, and count discrepancy: Full text pp. 13327-13328, section 4.2.2, Figure 2 and Table 4; Table 1 total 2,932 versus Table 4 positive total 2,928
  • Models and five experimental configurations: Full text pp. 13328-13329, sections 5.1-5.3
  • Automatic metrics and P-ACC classifier: Full text p. 13329, section 5.4
  • BLEU, ROUGE, perplexity, and P-ACC results: Full text pp. 13329-13330, section 5.5 and Table 5
  • Recognized linguistic limitations: Full text p. 13330, Discussion and Conclusion: English-like direct-translation expressions and future Korean profile source
  • Audited official repository: Official https://github.com/jiSilverH/psydial main, commit 08cd9a2c8037d65adf4b5ec4e77b3c072ed2982b; checked 15 Jul 2026
  • Label leakage in personality filtering: Official repository main_pipeline.py: dial_prompt and dial_response are appended to previous_questions_and_answers before tag_prompt_1 and tag_prompt_2 are sent
  • Profile filter does not observe the dialogue: Official repository main_filter.py: profile filter is prompted from the selected profile and expected pole rather than the generated dialogue
  • Generator reproducibility: Official repository main_pipeline.py: model alias gpt-4, random.choice without seed, legacy ChatCompletion API, temperature 0.5, max_tokens 1024 and presence_penalty 0.6
  • Moderation not connected: Official repository main_pipeline.py defines get_moderation but no call is present in the published pipeline
  • Missing artifacts and documentation: Official repository tree contains three scripts, overview image, 5_personality.csv, personaChat_personality_origin.csv and README; no final dialogue corpus, model/evaluation artifacts, dependency manifest or license; README names differ from two actual scripts
  • Comprehensive visual inspection: All 11 PDF pages rendered and visually inspected, including five tables, two figures, prompts, equations, dataset analysis, experimental settings, results, discussion, conclusion and references; checked 15 Jul 2026