Creating Multilingual Mental Health Dialogue Datasets: Limits of Persona-Based Localization via Nationality and Language

Applications, bias, and safety2026arXivApproved editorial review

Authors: Yunkai Xu, Saeed Abdullah

Keywords: Multilingual mental health, Synthetic personas, Clinical validity, LLM-as-a-judge, BDI-II, Cross-lingual evaluation, Dataset quality, Reproducibility

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

2
Authors
10
Findings
18
Limitations
6
Evidence

Editorial summary

English

This CLPsych 2026 workshop preprint asks a useful and deliberately narrow question: whether twelve English synthetic clinical personas preserve comparable depression-severity signals when only nationality and language are changed. The authors start from twelve prior profiles with distinct BDI-II totals and four highlighted symptoms. The remainder of each profile stays in English while four conditions are created: United States-English, China-Mandarin, Bangladesh-Bengali, and India-Hindi. ChatGPT-4o-mini generates five conversations per persona, after which five LLMs, ChatGPT-4o-mini, DeepSeek-v3.2, DeepSeek-R1-8B, Llama 3.1-8B, and Qwen3-8B, blindly compare all 66 persona pairs per language and choose which patient appears more depressed or return a tie. English has the highest reported accuracy for every judge. Published values range from 86.36 to 95.45 for GPT-4o-mini and from 83.33 to 93.75 for DeepSeek-v3.2, while open 8B models fluctuate more: DeepSeek-R1-8B moves from 63.33 in Bengali to 98.21 in English and Llama3-8B from 45.45 to 84.13. The paper interprets this pattern as evidence that changing two labels does not automatically produce equivalent clinical personas and recommends language- and context-specific output validation. That cautious conclusion is compatible with the design, but the results cannot separate language, culture, generator quality, and judge reliability: each country is coupled to one language, one call authors both voices, and no clinician or native-speaker review is included. Auditing the public repository exposes a more fundamental limitation. pipeline.py does not run two interacting agents; it makes one ChatGPT-4o-mini call and asks the model to write all five complete conversations, including therapist and patient turns. That call receives the exact BDI-II total, severity label, four symptoms, full profile, and scored definitions of all 21 BDI-II items. The task therefore measures how well another LLM recovers deliberately planted cues from one generator's synthetic text, not independent clinical inference or an emergent interview. The manuscript and artifact also diverge: the paper says five to seven turns per session and four symptom-specific sessions plus one mixed session, whereas the public prompt requests five examples averaging ten turns and does not encode that structure. Corpus quality prevents treating the release as a validated dataset. Bengali Elena-P2 contains a mixed Chinese-English phrase and a sleep response with Indonesian tidur followed by the Arabic/Urdu token خواب repeated 160 times; Maria-P9 includes _worker and النوم. These are not merely natural code-switches. More seriously, English Elena-P2 and Maria-P9 have SequenceMatcher similarity 0.99904: except for the initial name and punctuation in five headings, their conversations are identical, and every Elena dialogue turn speaks as Maria. Two profiles with hidden totals 35 and 40 thus receive effectively the same transcript, one mislabeled. Lower Bengali accuracy can reflect overt generator corruption, and English is not a clean baseline. The metrics also require care. Accuracy removes ties from its denominator. DeepSeek-R1-8B English has 55 correct decisions, one error, and ten ties: 55/56 yields the published 98.21%, but 55/66 is 83.33% when abstention counts toward total task performance. DeepSeek-v3.2 English has 60 correct, four wrong, and two ties, giving 93.75% without ties and 90.91% over all 66 pairs. The Llama3-8B accuracy and tie tables are arithmetically incompatible: 84.13% in English requires at least three ties and 65.38% in Hindi at least fourteen, already seventeen against the reported fourteen total before Bengali or Mandarin. Raw outputs are needed to resolve this. The so-called Same-Level Error Rate is not an error rate over pairs but the fraction of errors occurring within a category. DeepSeek-R1-8B's English 100% means its single non-tie error was within-level; it is not a stable estimate of calibration. The paper also labels BDI-II totals 12 and 13 as mild. Standard BDI-II ranges are 0-13 minimal, 14-19 mild, 20-28 moderate, and 29-63 severe, so P3 and P4 are mislabeled. Correcting them changes the number of within-category pairs from twelve to sixteen and alters the metric's classification. BDI-II measures self-reported symptom severity and is not diagnostic by itself; here the ground truth is a hidden prompt total, not an independent inventory administered from the dialogue. Reproducibility is partial. The repository releases 48 localized profiles and 48 chat files, the prompt, hidden totals, and two judge scripts, and all Python compiles. But the generator is hard-coded to P7 and Bengali only; other languages are commented and English is absent. It uses mutable gpt-4o-mini without a snapshot, seed, temperature, or request metadata. The API judge configures only deepseek-chat at temperature 1.3; the local script only deepseek-r1:8b and an invalid URL. There are no configurations for the other three judges, result JSONL/CSV files, aggregator, metric or table code, dependency specification, lockfile, tests, CI, license, or runnable instructions. The corpus can be inspected but the reported tables cannot be regenerated. The paper itself appropriately acknowledges no significance tests, no human validators, and no separation of generation from judging, and warns against diagnosis or deployment. That caution must remain. An opening sentence in the Ethics section about trauma, prolonged-exposure therapy, psychotherapists, and future license restrictions should not be treated as evidence: it is unrelated pasted material and conflicts with the unlicensed repository. The defensible contribution is not a clinically valid multilingual dataset or a general proof that localized personas fail. It is a valuable negative case study: shallow substitution does not guarantee equivalence, LLM-as-judge evaluation can amplify generator failures, and mental-health corpora require human linguistic QA, correct clinical categories, raw outputs, and abstention-aware metrics before cross-language comparisons are credible.

Español

Este preprint aceptado en el workshop CLPsych 2026 estudia una pregunta útil y deliberadamente limitada: si doce personas clínicas sintéticas creadas en inglés conservan señales comparables de gravedad depresiva cuando solo se cambian nacionalidad e idioma. Los autores parten de doce perfiles previos con puntuaciones BDI-II distintas y cuatro síntomas destacados. Mantienen en inglés el resto del perfil y producen cuatro condiciones: Estados Unidos-inglés, China-mandarín, Bangladesh-bengalí e India-hindi. ChatGPT-4o-mini genera cinco conversaciones por persona y después cinco LLM, ChatGPT-4o-mini, DeepSeek-v3.2, DeepSeek-R1-8B, Llama 3.1-8B y Qwen3-8B, comparan a ciegas los 66 pares posibles de cada idioma para decidir qué paciente parece más deprimido o declarar empate. La exactitud inglesa es la mayor para todos los jueces. Los porcentajes publicados van de 86,36 a 95,45 para GPT-4o-mini, de 83,33 a 93,75 para DeepSeek-v3.2 y fluctúan más en los modelos abiertos de 8B; DeepSeek-R1-8B pasa de 63,33 en bengalí a 98,21 en inglés y Llama3-8B de 45,45 a 84,13. El artículo interpreta que sustituir dos etiquetas no crea automáticamente personas clínicas equivalentes y recomienda validar cada artefacto multilingüe en su idioma y contexto. Esa conclusión prudente está alineada con el diseño, pero los resultados no permiten separar idioma, cultura, calidad del generador y fiabilidad del juez: cada país está acoplado a un solo idioma, una sola llamada genera ambas voces y no hay revisión de clínicos ni hablantes nativos. La auditoría del repositorio público hace visible una limitación más fundamental. pipeline.py no ejecuta dos agentes interactivos: realiza una única llamada a ChatGPT-4o-mini y le pide escribir completas las cinco conversaciones, tanto terapeuta como paciente. Esa llamada recibe la puntuación BDI-II exacta, la etiqueta de gravedad, los cuatro síntomas, el perfil íntegro y la definición puntuable de los 21 ítems. La evaluación mide, por tanto, cuánto recupera otro LLM unas pistas insertadas deliberadamente por un generador, no una inferencia clínica independiente ni una entrevista emergente. También hay una divergencia entre manuscrito y artefacto: el texto dice cinco a siete turnos por sesión y cuatro sesiones por síntoma más una mixta, mientras el prompt público pide cinco ejemplos de unas diez intervenciones y no codifica esa estructura. La calidad del corpus impide tratarlo como dataset validado. Elena-P2 en bengalí contiene una frase chino-inglesa y una respuesta sobre el sueño con la palabra indonesia tidur seguida del token árabe/urdu خواب repetido 160 veces; Maria-P9 contiene _worker y النوم. El problema no es solo code-switching natural. Más grave, los chats ingleses Elena-P2 y Maria-P9 tienen similitud 0,99904: salvo el nombre inicial y la puntuación de cinco encabezados son idénticos, y todos los turnos de Elena hablan como Maria. Así, dos perfiles con puntuaciones ocultas 35 y 40 reciben prácticamente el mismo texto, uno mal etiquetado. La menor exactitud bengalí puede reflejar corrupción evidente del generador, y el inglés no es una línea base limpia. La métrica también exige cautela. La exactitud elimina los empates del denominador. En DeepSeek-R1-8B inglés hay 55 aciertos, un error y diez empates: 55/56 produce el 98,21% publicado, pero 55/66 es 83,33% si las abstenciones cuentan en el rendimiento total. En DeepSeek-v3.2 inglés, 60 aciertos, cuatro errores y dos empates equivalen a 93,75% sin empates y 90,91% sobre los 66 pares. Además, la tabla de exactitud y la tabla de empates de Llama3-8B son aritméticamente incompatibles: 84,13% en inglés requiere al menos tres empates y 65,38% en hindi al menos catorce, ya diecisiete frente a los catorce totales declarados antes de considerar los otros idiomas. Sin resultados crudos no puede resolverse. La llamada Same-Level Error Rate tampoco es una tasa sobre pares, sino el porcentaje de los errores que ocurre dentro de una categoría. El 100% inglés de DeepSeek-R1-8B significa que su único error no-empate fue intranivel; no estima de forma estable una calibración. Encima, el artículo clasifica 12 y 13 como depresión leve. Los rangos estándar de BDI-II son 0-13 mínimo, 14-19 leve, 20-28 moderado y 29-63 grave, por lo que P3 y P4 están mal etiquetados. Esto cambia de doce a dieciséis los pares intranivel y altera el numerador del indicador. BDI-II mide gravedad autoinformada y no diagnostica por sí solo; aquí el ground truth es una puntuación oculta del prompt, no una administración independiente del inventario sobre el diálogo. La reproducibilidad es parcial. El repositorio libera 48 perfiles localizados y 48 archivos de chat, el prompt, los totales ocultos y dos jueces, y todo el Python compila. Pero el generador está fijado a P7 y solo bengalí; los otros idiomas están comentados y el inglés no existe en la configuración. Usa el alias mutable gpt-4o-mini sin snapshot, semilla, temperatura ni metadatos de llamada. El juez API solo configura deepseek-chat a temperatura 1,3; el local solo deepseek-r1:8b y tiene una URL inválida. No hay configuraciones de los otros tres jueces, resultados JSONL/CSV, agregador, código de métricas o tablas, dependencias, lockfile, tests, CI, licencia ni instrucciones ejecutables. Por ello puede auditarse el corpus, pero no regenerar las tablas. El propio artículo reconoce que no hay pruebas de significación, revisores humanos ni separación entre generación y evaluación, y advierte contra diagnóstico o despliegue. Esa cautela debe conservarse. También debe ignorarse como evidencia una primera frase de la sección Ethics que habla de trauma, terapia de exposición prolongada, psicoterapeutas y futuras restricciones de licencia: es material ajeno pegado en el manuscrito y contradice el repositorio sin licencia. La contribución defendible no es un dataset clínico multilingüe ni una demostración general de que las personas localizadas fracasan. Es un caso negativo valioso: la sustitución superficial no garantiza equivalencia, la evaluación LLM-as-judge puede amplificar errores del generador y cualquier corpus de salud mental requiere QA lingüístico humano, categorías clínicas correctas, datos crudos y métricas que contabilicen abstenciones antes de sostener comparaciones entre idiomas.

Research question

Do twelve synthetic depression personas constructed in English retain comparable BDI-II severity signals when only nationality and language are changed to Mandarin, Bengali, and Hindi, and with what consistency do different LLM judges rank them?

Method

Adapts twelve previous personas with unique BDI-II scores to four country-language pairs, generates five complete conversations per persona with a single call to ChatGPT-4o-mini, and asks five LLM judges to compare the 66 pairs of each language. Reports accuracy conditioned on non-ties, proportion of intra-level errors, and distance of ties. The audit recalculates denominators, checks BDI-II categories, inspects the 48 chats, and reviews the public code.

Sample: Twelve synthetic profiles, four linguistic-national conditions, and 66 comparisons per condition and judge. There are no real participants, clinicians, native speakers, or human annotators. A single generation per profile and a single judgment per model and pair; raw evaluation outputs are not published.

Findings

  • English obtains the highest published accuracy for the five judges; Bengali and Hindi are the weakest scenarios in several 8B models.
  • The cautious conclusion that substituting language and nationality does not guarantee equivalence is supported as a case study.
  • Accuracy excludes ties and can inflate overall performance: DeepSeek-R1-8B English drops from 98.21% to 83.33% over the 66 pairs.
  • The accuracy and tie tables for Llama3-8B are arithmetically incompatible.
  • The Same-Level Error Rate indicator conditions on errors and can be based on a single case.
  • P3 and P4, with BDI-II 12 and 13, are misclassified as mild; they belong to the standard minimal range.
  • The single generator receives the BDI-II score and the definitions of its 21 items and writes both voices, so the test recovers inserted signals.
  • Elena-P2 Bengali contains severe multi-script corruption, including a 160-token Arabic/Urdu repetition.
  • Elena-P2 and Maria-P9 English are practically identical, and the Elena file speaks as Maria.
  • The code compiles, but the repository does not reproduce the tables and does not fix versions or stochastic settings.

Limitations

  • Nationality and language are coupled one to one; the effect of neither is identified separately.
  • There is no validation by clinicians, native speakers, or bilingual annotators.
  • Naturalness, clinical realism, symptom fidelity, cultural adequacy, safety, or therapeutic quality are not evaluated.
  • A single call generates therapist and patient and knows the hidden ground truth.
  • The ground truth is a prompt score, not a BDI-II independently administered to the dialogue.
  • The twelve totals are unique and force ordering one-point differences even when the text does not cover the 21 items.
  • There are no generation or judgment repetitions, seeds, intervals, significance, or a model of dependence between pairs.
  • Accuracy discards abstentions; comparing judges with different propensity to tie is not straightforward.
  • Same-Level Error Rate is poorly named, has small denominators, and uses two incorrect categories.
  • A central table cannot be arithmetically reconciled with the tie table.
  • The corpus contains corrupt Bengali text and a mislabeled English duplicate.
  • The public prompt contradicts the length and structure described in the method.
  • The released pipeline only runs P7 in Bengali without editing the code.
  • Model aliases are mutable and lack snapshot, temperature, seed, and generation metadata.
  • Only two of five judges have partial configuration and one local URL is invalid.
  • No results, failure logs, aggregation code, metrics, or tables are released.
  • Dependencies, lockfile, tests, CI, license, instructions, and sensitive-use governance are missing.
  • The Ethics section includes unrelated material about another dataset and another therapy.

What the study does not establish

  • It does not demonstrate that all localization of synthetic personas fails.
  • It does not demonstrate that the observed differences are cultural rather than linguistic or technical.
  • It does not validate a multilingual clinical dataset suitable for training or deployment.
  • It does not validate diagnosis, screening, therapy, or replacement of professionals.
  • It does not show that LLM judges are correct without independent human reference.
  • It does not prove that English is a clean or clinically validated baseline.
  • It does not evaluate a real interaction between a therapist agent and a patient agent.
  • It does not allow clean comparison of models when ties are excluded.
  • It does not reproduce the published tables from the public artifact.
  • It does not causally attribute the worst performance to low resource, language, nationality, or culture.

Traceability

Scope: Full text

Version: arXiv:2606.19640v1

Consulted source: https://arxiv.org/abs/2606.19640v1

Review: Codex 15-page full-text visual, complete TeX, corpus, clinical-cutoff, metric-arithmetic, code and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • ChatGPT-4o-mini as dialogue generator and judge
  • DeepSeek-v3.2
  • DeepSeek-R1-8B
  • Llama 3.1-8B
  • Qwen3-8B

Instruments and metrics

  • BDI-II hidden persona totals and severity categories
  • Pairwise higher-depression judgment
  • Non-tie conditional accuracy
  • Same-Level Error Rate
  • Tie count and BDI-II score distance

Data used

  • Twelve English baseline depression personas from Wang et al.
  • Forty-eight localized persona files
  • Forty-eight generated chat artifacts, intended as 60 sessions per language
  • Public CLPsych2026workshop GitHub repository at commit f05565762dd290a5823d842995e69d359ecb3bde

Evidence and location

  • Metadata, version, authors, and CLPsych acceptance: Official arXiv record 2606.19640v1, checked 2026-07-16
  • Construction, generation, judges, metrics, and results: arXiv v1, Method, Findings, Tables 1-4 and Appendix
  • Declared limitations and clinical warnings: arXiv v1, Discussion, Limitations and Ethical Considerations
  • Prompt, configuration, corrupt corpus, duplicate, and reproduction gaps: Public repository Xuyk021/CLPsych2026workshop at commit f05565762dd290a5823d842995e69d359ecb3bde
  • Standard BDI-II ranges and diagnostic threshold: AHRQ NCBI Bookshelf, Definition of Treatment-Resistant Depression, Table 8; Pearson BDI-II product documentation
  • Reconstruction of denominators, Llama inconsistency, and comprehensive audit: reports/verification/article-296-multilingual-mental-health-dialogue-corpus-label-metric-arithmetic-code-and-validity-audit.json