Multilingual Transformer-Based Medical Chatbots in Low-Resource Dialectal Settings

Introduction: Developing medical chatbots that are proficient in under-resourced languages and variants, such as Algerian Arabic, which blends Arabic and French and does not have standardized linguistic resources, is currently an area of great interest and challenge. Objective: This paper proposes a multilingual medical question answering system with capabilities to process and respond to Algerian Arabic and French to bridge the resource gap in healthcare-oriented dialogue systems. Methods: A large set of more than 74k medical QA pairs was obtained from social media sites and medical QA resources, followed by strict preprocessing and fine-tuning of a set of transformer architectures such as GPT-2, mT5-small, as well as a hybrid DistilXLM-R-DistilGPT2 model. Results and Discussion: The performance for the mT5-small model yielded the optimum fluency and accuracy (F1 value ≈ 47%, andBLEU value ≈ 40%), proving the effectiveness of the encoder-decoder model in the multilingual medical conversation task. GPT-2 displayed moderate fluencies and lower precision, while the hybrid model displayed higher recalls and lower coherence. This suggests that with higher-quality datasets and ethical validation, the model has the potential to lead to more accurate and easily accessible medical chatbots in low-resource settings for quality health care. Conclusion:: Model architecture has a remarkable impact on multilingual medical dialogue performance in under-resourced settings. In particular, encoder-decoder models, such as mT5, show great promise for handling dialectal Arabic and French, setting a valuable benchmark for future research and development of healthcare chatbots.