Abstract: Sunny Rai, Khushi Shelat, Devansh Jain, Ashwin Kishen, Young Min Cho, Maitreyi Redkar, Samindara Hardikar-Sawant, Lyle Ungar, Sharath Chandra Guntuku. Proceedings of the 3rd Workshop on Cross-Cultural Considerations in NLP (C3NLP 2025).
2025.
Introduction
Current Natural Language Processing (NLP) models for mental health are predominantly trained on "WEIRD" (Western, Educated, Industrialized, Rich, and Democratic) data. This geographic and cultural concentration creates a significant bias, as mental health is not a universal language. Cultural norms deeply influence how individuals perceive, experience, and articulate psychological distress.
This "West-centric" bias in AI development leads to systemic health inequity. For instance, models trained on American English often prioritize "I-talk" (first-person singular pronouns) as a primary marker for depression. However, individuals from collectivist cultures, such as India, may express distress through different linguistic markers—focusing on family obligations or seeking direct advice rather than self-reflection. This research quantifies these cross-cultural differences to facilitate the creation of more culturally competent mental health models.
Methodology
The researchers employed a multi-pronged computational approach to compare language patterns between Indian users and the "Rest of the World" (RoW):
- Feature Extraction: Three types of interpretable features were utilized: N-grams (1-3 words) to capture cultural idioms of distress; LIWC-22 (Linguistic Inquiry and Word Count) to categorize words into psychosocial groups; and LDA (Latent Dirichlet Allocation) for unsupervised topic modeling.
- Correlation Analysis: Pearson correlation and ordinary least squares regression identified linguistic features most strongly associated with each cohort.
- Predictive Modeling: "One-vs-rest" logistic regression models were trained to determine if these features could accurately distinguish between cultural groups.
- Clinical Validation: The authors collaborated with two clinical psychologists practicing in India to validate whether the data-driven topics aligned with real-world clinical experiences in the Indian context.
Dataset and Cohort Construction
The study utilized a massive scrape of Reddit data via the PushShift API. Users were geolocated using the "GLOBAL" inference model and categorized into four distinct cohorts:
- MH-India: Mental health-related posts from India.
- MH-RoW: Mental health-related posts from the Rest of the World.
- Control Groups: Regional users posting in non-mental health forums.
After filtering for post length and geolocation accuracy, the final dataset included 4,530 distinct users and nearly 300,000 posts.
Results and Key Insights
Linguistic Markers of Distress
Analysis revealed that Indian users are significantly more likely to use "help-seeking" language (e.g., "suggest," "please help") compared to the "support-seeking" language common in Western cohorts, which focuses on sharing internal feelings. Furthermore, mental health discussions in the Indian cohort are inextricably linked to academic and career pressures (e.g., "college," "exam," "study"), whereas RoW users utilize more clinical terminology (e.g., "mental," "health," "diagnosis").
Psychosocial Category Differences
A stark divide exists in social orientation. Western mental health expressions are heavily characterized by first-person singular pronouns (I, me, my). In contrast, Indian expressions show a higher correlation with second and third-person pronouns (you, we), reflecting a collectivist social orientation even during personal distress. Additionally, Indian users exhibit a "present focus" on current struggles, while RoW users are more "past-focused," ruminating on personal histories.
Model Performance and Validation
The logistic regression models achieved high AUC scores (0.758 to 0.853) for the MH-India group, proving that the linguistic signature of mental health in India is distinct and machine-recognizable. Clinical validation further supported these findings, with psychologists agreeing with machine-generated topics 81.49% of the time and confirming that 95% of the top 20 India-associated topics were prevalent in their clinical practice.
Discussion and Limitations
While the study provides a robust framework for cultural NLP, it faces several limitations. The "Reddit Bias" means the sample is limited to English-speaking, tech-savvy Indians, which does not represent the full diversity of the subcontinent. Additionally, text-based geolocation inference is not 100% accurate and may include expats.
Future AI interventions must be "precision-targeted." For example, a mental health chatbot for Indian users should prioritize providing actionable advice and addressing academic/familial pressures rather than solely reflecting emotions. Moving forward, expanding this research to regional languages and rural populations is essential for true health equity.
Summary updated