"Kya family planning after marriage hoti hai?": Integrating Cultural Sensitivity in an LLM Chatbot for Reproductive Health

Research Background and Issues

• Issues or Challenges:
  The authors identified significant challenges faced by many global communities, particularly urban poor communities in India, in accessing sexual and reproductive health information. These challenges include cultural taboos, limited healthcare resources, and restrictive social and legal norms. Furthermore, existing large language models (LLMs), which are predominantly trained on English, often fail to capture local dialects, cultural contexts, and taboos, resulting in a lack of resonance with populations in remote areas.

• Significance:
  Sexual and reproductive health is a cornerstone for addressing physical, psychological, and familial balance, directly impacting individual and community well-being. Integrating technology into this domain can bridge gaps in healthcare resources and simplify the dissemination of health education. However, culturally insensitive technological implementations may exacerbate information inequality.

• Research Motivation and Related Work:
  The authors aim to design a culturally sensitive LLM chatbot centered on "urban low-income communities in India," providing women with unbiased and personalized sexual and reproductive health information. The study draws on theories of culturally sensitive health design, incorporating the unique contexts of Indian communities, such as religious traditions, gender roles, and social dynamics. Additionally, the paper references existing work on health chatbots (e.g., research on breastfeeding education), which highlights the importance of cultural factors in user trust and engagement.

Solution

• Method or Solution:
  The authors collaborated with a Mumbai-based NGO, “Myna Mahila Foundation,” to develop a GPT-4-based LLM chatbot. The chatbot aims to improve urban low-income women’s access to health information by answering questions about family planning, contraception, and health-related topics.

• Innovations:

  1. Multilingual Support: The chatbot supports “Hinglish” (Hindi written in Roman script) and English to cater to the preferred language forms of the target audience.
  2. Culturally Sensitive Design: The chatbot employs predefined prompts to mimic the tone of Indian female obstetricians, integrating religious, medical ethics, and diverse user interaction styles.
  3. Technical Architecture: By combining translation services, retrieval-augmented generation (RAG) technology, and custom dictionaries, the system ensures accurate comprehension of user queries and generates responses aligned with cultural contexts and medical science.
  4. Community Collaboration Approach: Knowledge bases and system testing were co-developed with participants from the target community to ensure information aligns with local cultural and linguistic styles.

• Implementation Steps and Key Technologies:

  1. Translation Module: Converts user input in Hinglish or text with spelling errors into structured English queries.
  2. Retrieval Module: Searches embedded medical documents to retrieve relevant background information and reduce misinformation.
  3. Localization Module: Adjusts response text using language and terminology familiar to users while ensuring simplicity and clarity.
  4. User Data Analysis and Iterative Optimization: Analyzes 2,128 interaction logs and focus group feedback to continuously refine system design.

Research Outcomes

• Specific Results:

  1. The chatbot successfully answered 2,118 queries, most of which revolved around family planning, contraception, and reproductive health.
  2. Users expressed trust in the chatbot’s design, particularly its neutral and non-judgmental responses to sensitive topics (e.g., gender selection), which encouraged open public discussions.
  3. The system effectively addressed complex social dynamics (e.g., the influence of family pressure and religious beliefs on health decisions) and provided personalized health advice rooted in community culture.

• Advantages Compared to Existing Solutions:

  1. Better capture of cultural and linguistic nuances, such as local dialects and user preferences.
  2. Improved balance between medical advice and social dynamics by engaging with religious leaders or community representatives to enhance consumer trust.
  3. The use of dictionary updates and RAG technology mitigates issues of overfitting or semantic misunderstandings common in traditional language models.

• Experimental or Evaluation Results:
  Experimental results showed that participants demonstrated more efficient information acquisition and engagement after using the chatbot. However, the chatbot still faced limitations in accurately interpreting semantic ambiguities, certain social norms, or cultural taboos.

• Limitations and Future Directions:

  1. The inability to fully link user IDs or identity background data made it difficult to evaluate the effectiveness of personalized recommendations.
  2. Temporary cultural shifts (e.g., subtle language variations) or complex religious conflicts were not fully captured.
  3. Future work could expand to broader geographic and social groups and optimize voice-activated services for users with low digital literacy. Additionally, the system could enhance recommendations regarding the cost or availability of assisted reproductive technologies (e.g., IVF).