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 to enhance patient care and reduce costs.
• Existing medical dialogue systems often struggle with knowledge graph incompleteness, limiting their ability to use dialogue history for accurate responses.

Purpose of the Study:

• To propose a general method for embedding knowledge graph triples into large-scale models to generate clinically correct responses using dialogue history.
• To address the limitations of current knowledge-grounded dialogue models in handling incomplete knowledge graphs.

Main Methods:

• Developed a Masked Entity Dialogue (MED) model that embeds knowledge graph triples into scalable models.
• Implemented a masked entity prediction approach using cross-entropy loss to learn contextual information from dialogue history.
• Fine-tuned the MED model on the MedDialog(EN) and Covid Dataset, and utilized a Medical Entity Prediction (MEP) model for knowledge graph augmentation.

Main Results:

• The proposed MED model effectively incorporates dialogue history and knowledge graph information to generate contextually relevant and clinically accurate responses.
• Empirical results on MedDialog(EN) and Covid Dataset show superior performance compared to state-of-the-art methods in both automatic and human evaluations.
• The MEP model successfully augmented knowledge graphs with data-specific medical information.

Conclusions:

• The novel approach significantly improves the performance of knowledge-grounded medical dialogue systems.
• The MED model offers a promising solution for generating more accurate and human-like medical conversations by effectively utilizing dialogue context and knowledge graphs.
• The study highlights the importance of addressing knowledge graph incompleteness and enhancing medical knowledge bases for advanced healthcare AI.