Culturally-aware support

Abstract

Large language models (LLMs) have shown promise in providing emotional support and generating empathetic responses for individuals in distress. However, their ability to deliver culturally sensitive emotional support remains underexplored. In this work, we introduce CultureCare, the first dataset designed specifically for this task. It spans four distinct cultures and includes fine-grained annotations that capture both emotional and cultural dimensions. Leveraging CultureCare, we evaluate state-of-the-art LLMs across several adaptation strategies through both automatic and human evaluations. Our results show that adapted LLMs outperform anonymous online peer support and that including simple matching cultural information in the prompt is insufficient for delivering culturally sensitive support. We further explore the potential application of LLMs in clinical training settings for cross-cultural therapy, as experts emphasize their value in training psychology students in culturally sensitive practice.

We present the CultureCare dataset!

CultureCare is a fine-grained labelled multi-cultural dataset for emotional distress and support.💬
It's designed to evaluate and adapt large language models (LLMs) in generating culturally-aware emotional support ❤️.

🤝 Empowering Peer Support

By helping AI understand cultural nuances in emotional expression, CultureCare enables the creation of empathetic peer support systems that are inclusive, sensitive, and deeply human-centered.

This is vital in communities where access to trained professionals is limited or where stigma prevents people from seeking formal help.

🎓 Training Culturally Competent Therapists

The dataset also serves as a resource to train young psychologists and counselors in recognizing cultural contexts behind distress signals.

By exposing therapists-in-training to diverse narratives and emotional cues, CultureCare promotes culturally competent therapy, helping close the empathy gap in global mental health.

🌐 A Step Toward Inclusive AI

CultureCare paves the way for more equitable NLP applications—where language models can engage respectfully and supportively across cultures.

Dataset details

Construction pipeline

Our Data Collection and Annotation Process:

(1) We collect data by querying culture-specific subreddits with mental health related keywords, and querying mental health subreddits with culture-specific keywords.

(2) To remove noisy instances, we apply rule-based and LLM-based filtering methods to ensure data quality.

(3) In-culture annotators (i.e., annotators whose culture matches the data’s culture) label posts and responses for:

• Emotional distress messages and their intensity,
• Cultural signal and their types,
• Support messages and their strategies.

An additional in-culture annotator then verifies the correctness and completeness of all annotations.

🗂️ A new dataset: 462 instances, 4 cultures, fine-grained annotations.

Examples

Cultural themes (mined from the annotated cultural signals)

Adapting LLMs for culturally-aware emotional support

We experiment with four key adaptation strategies for prompting large language models (LLMs) to generate culturally-aware responses: Standard, which prompts the model to behave like a Redditor matching the data context as a baseline; Culture-Informed Role-Playing (+culture), instructing the LLM to role-play the cultural background to foster empathetic, culturally aligned responses; Guided Principles / Constitutions (+guided), which provides LLMs with culturally sensitive counseling guidelines based on the CCCI-R framework to emulate professional cross-cultural competency; and Explicit Cultural Signals (+annotation), where distress message annotations from the dataset are explicitly included in the prompts to enrich contextual understanding. Finally, a Combined (+cga) strategy merges all three methods—culture role-play, guided principles, and annotations—while refining guidelines to remove redundant cross-cultural items, enabling LLMs to receive comprehensive cultural context and instructions for more nuanced and appropriate support.

Clinical psychologists say this resource helps train culturally competent therapists!

BibTeX

@article{liu2025tailoredemotionalllmsupporterenhancing,
      title={Tailored Emotional LLM-Supporter: Enhancing Cultural Sensitivity},
      author={Chen Cecilia Liu and Hiba Arnaout and Nils Kovačić and Dana Atzil-Slonim and Iryna Gurevych},
      year={2025},
      note={Chen Cecilia Liu and Hiba Arnaout contributed equally to this work},
      journal={ArXiv preprint},
      year={2025},
      url={https://arxiv.org/abs/2508.07902},
      doi={https://doi.org/10.48550/arXiv.2508.07902},
      eprinttype={arXiv},
      eprint={2508.07902}
}