Understanding Patient Query With Weak Supervision From Doctor Response

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• The demand for automated diagnostic dialogue systems is growing.
• Slot filling is crucial for converting medical queries into structured data for diagnosis.
• High-quality datasets are scarce, hindering the performance of slot-filling models.

Purpose of the Study:

• To propose a two-step training framework to effectively utilize unlabeled medical dialogue data.
• To address the limitations of existing datasets for slot filling in diagnostic systems.
• To enhance the performance of slot filling in medical dialogue systems.

Main Methods:

• A two-step training framework designed to leverage unlabeled dialogue data from medical communities.
• Utilizing colloquial input and professional responses from diagnostic dialogues.
• Development and release of a Chinese dataset with annotated and unlabeled samples.

Main Results:

• The proposed method demonstrated significant improvements over strong baselines.
• Achieved an average increase of 6.32% in Micro F1 and 8.20% in Macro F1.
• Experimental results validate the effectiveness of the two-step training framework.

Conclusions:

• The proposed framework effectively utilizes unlabeled medical dialogue data to improve slot filling.
• The developed Chinese dataset facilitates further research in diagnostic dialogue systems.
• This approach offers a promising solution for enhancing automatic diagnosis through improved slot filling.