Knowledge grounded medical dialogue generation using augmented graphs.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Smart healthcare systems leverage health data for improved access, reduced costs, and high-quality care.
• Existing medical dialogue systems use pre-trained models and knowledge bases like UMLS but struggle with knowledge graph incompleteness.
• Current models fail to integrate dialogue history into entity embeddings, degrading performance.

Purpose of the Study:

• To propose a general method for embedding knowledge graph triples into large-scale models for clinically correct response generation.
• To address the limitations of knowledge graph incompleteness in dialogue systems.
• To enhance medical dialogue systems' ability to utilize conversation history.

Main Methods:

• Developed a Masked Entity Dialogue (MED) model to embed knowledge graph triples.
• Implemented a masked entity prediction approach using cross-entropy loss on triples overlapping with patient utterances.
• Fine-tuned the MED model on the Covid Dataset and re-curated/augmented knowledge graphs using a Medical Entity Prediction (MEP) model.

Main Results:

• The proposed model effectively learns contextual information from dialogues for clinically correct responses.
• Empirical results on MedDialog(EN) and Covid Dataset show superior performance compared to state-of-the-art methods.
• Both automatic and human evaluations confirmed the model's effectiveness.

Conclusions:

• The novel approach enhances knowledge-grounded medical dialogue systems by integrating dialogue history and improving knowledge graph representations.
• The MED and MEP models offer a significant advancement in generating accurate and contextually relevant medical responses.
• This research contributes to the development of more sophisticated and reliable smart healthcare systems.