Xingyu LIU

Access to information on women's intimate health remains a critical issue, particularly in a context where certain topics are still taboo and medical consultations are often time-constrained. This thesis is part of the development of an online personalized counseling solution and proposes the design of a dialogue support system intended for healthcare professionals, with the goal of improving communication with patients in the context of teleconsulting.

 Unlike traditional dialogue systems that generate responses directly for users, our approach acts as a mediator, assisting midwives and gynecologists in formulating appropriate responses, while preserving their freedom to modify or reject the responses.
 

The system architecture is based on a hybrid model, combining NLU model, a prototype-based response retrieval module with a generative Large Language Model (LLM). Training the models required building a domain-specific dataset by extracting and adapting a corpus in English, as well as generating artificial dialogues using generative models. An annotation process structured the information of women's intimate health, enabling better adaptation to the healthcare domain.
 

The system was evaluated through an experimental study involving midwives, domain experts and naive participants, measuring usability, response quality, and efficiency in time management. Results indicate a response time reduction of 36.20 to 75.40 seconds for healthcare professionals, along with satisfying performance in coherence (3.7/5) and fluency (4.5/5) of generated responses. Additionally, the high semantic similarity (0.805) between midwives' responses when using the system highlights its potential to standardize medical communication.