Leveraging GraphRAG with Large Language Models to Identify Help-Seeking Information Among Domestic Violence Survivors...

Area of Science:

• Public Health
• Natural Language Processing
• Artificial Intelligence

Background:

• Domestic violence is a complex public health issue requiring in-depth analysis of survivor experiences.
• Existing topic modeling techniques may lack the nuance needed for culturally sensitive narratives.

Purpose of the Study:

• To develop and evaluate a hybrid framework for scalable and interpretable topic modeling of domestic violence survivor narratives.
• To improve the analysis of qualitative data in sensitive contexts.

Main Methods:

• A hybrid framework integrating Microsoft GraphRAG with GPT-4o-mini was proposed.
• The framework was applied to analyze transcribed interviews from domestic violence survivors.
• Performance was compared against established topic modeling methods like LDA, BERTopic, and TopicGPT.

Main Results:

• The GraphRAG framework achieved a coherence score of 0.79, outperforming other methods.
• It demonstrated 97% accuracy in identifying 70 distinct entities.
• The approach proved effective for culturally sensitive data analysis.

Conclusions:

• The hybrid GraphRAG and GPT-4o-mini framework offers a significant advancement in topic modeling for domestic violence survivor narratives.
• This method enhances interpretability and scalability, providing better support for survivors through improved data analysis.