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

Researchers reveal significant limitations in using English-centric persona-based methods to generate multilingual mental health datasets, finding that simply adding nationality and language parameters introduces clinical inconsistencies and causes LLM evaluators to perform poorly on non-English depression severity assessments. The study underscores the urgent need for culturally responsive data generation approaches to build equitable AI mental health systems globally.

This research addresses a critical gap in AI-driven mental health infrastructure that disproportionately affects non-English speaking populations. As LLMs increasingly power mental health support systems, the reliance on English-centric training data creates systemic biases that compromise clinical accuracy across languages and cultures. The study demonstrates that mechanical localization—simply translating personas or adding language parameters—fails to capture the nuanced cultural and clinical dimensions necessary for accurate mental health assessment.

The broader context reveals a pattern in AI development where English-language datasets dominate training and evaluation frameworks. Mental health terminology, symptom expression, and severity indicators vary significantly across cultures, making direct translation insufficient. When LLM judges evaluate depression severity in Mandarin, Bengali, and Hindi, their performance degradation exposes a fundamental architectural flaw: models trained predominantly on English data lack the cultural context needed for cross-linguistic clinical assessment.

For developers building mental health AI applications, this research signals the necessity of investing in culturally grounded data collection rather than relying on synthetic persona generation alone. Mental health platforms serving diverse populations face reputational and clinical risks if their assessment algorithms exhibit systematic inaccuracy in non-English contexts. This particularly impacts emerging markets where mental health AI adoption is accelerating but localized validation remains minimal.

The pathway forward requires collaborative approaches combining synthetic data generation with native-speaker validation, cultural consultation, and language-specific fine-tuning of evaluation models. Organizations developing mental health AI must prioritize culturally responsive methodologies from inception rather than treating localization as a post-deployment consideration.

• →English-centric persona-based methods introduce clinical inconsistencies when directly applied to multilingual mental health datasets without cultural adaptation.
• →LLM judge models demonstrate measurable performance degradation and inaccuracies when assessing depression severity in non-English languages.
• →Simple localization through nationality and language parameter modifications is insufficient for generating clinically consistent multilingual mental health data.
• →Culturally responsive data generation methodologies are essential for building equitable and accurate global mental health AI systems.
• →Mental health platform developers must invest in native-speaker validation and cultural consultation rather than relying solely on synthetic data localization.