Mai: A Transformer Based Domain Specific Chatbot for Menstrual Health

Abstract

Listen

Translate article

Traditional health-seeking behaviors, particularly in menstrual health, are deeply rooted in local languages, sociocultural norms, and religious beliefs. These factors significantly influence women in South Asia and other developing regions, often proving detrimental to their physical and psychological health. Recognizing the multifaceted nature of these challenges, this paper investigates the role of generative domain-specific chatbots as a supportive mechanism to enhance menstrual health-seeking behaviors among South Asian women. Specifically, it aims to create and assess Mai, a conversational agent, which utilizes Roman-Urdu and English models to provide accurate information, thereby contributing to the broader effort of empowering women with knowledge and resources. Due to a lack of conversational data in this domain, our research involves creating a dataset from existing English sources and translating it into Roman Urdu, presenting its own set of challenges. These datasets, specifically curated for this research, addressed a wide range of menstrual health topics. Despite being an English monolingual model, DialoGPT’s transformer architecture produced favorable results when fine-tuned on a Roman Urdu dataset. While the Roman Urdu version occasionally faced challenges with contextual understanding, both models achieved promising results. Automatic evaluation metrics indicated strong performance for Mai compared to baseline models. Likewise, user evaluations yielded average ratings of 4.2 and 4.3 (out of 5) for the Roman Urdu and English chatbots, respectively, and more than half of the medical professionals surveyed expressed willingness to recommend the dialogue system. Overall, our findings highlight the transformative potential of domain-specific dialogue systems in health education, with room for improvement in overcoming cultural and linguistic barriers to women’s health information access.