Predicting the Information Need for Domestic Violence Survivors Based on the Fine-Tuned Large Language Model

Area of Science:

• Digital Health
• Natural Language Processing
• Social Sciences

Background:

• Domestic violence survivors often avoid in-person help due to embarrassment.
• Online health communities provide a crucial avenue for survivors to seek emotional support.
• Timely understanding and response to online disclosures are vital for victim support.

Purpose of the Study:

• To develop and evaluate a fine-tuned large language model (LLM) for predicting informational needs in online posts from domestic violence survivors.
• To enhance the ability of healthcare providers to offer timely and relevant support.

Main Methods:

• Fine-tuning the Llama2-7B-chat model using a dataset of 273 Reddit posts manually annotated by domain experts.
• Developing guidance for identifying information needs within the posts.
• Evaluating model performance on a random sample of 15 posts.

Main Results:

• The fine-tuned LLM demonstrated an accuracy of 66.6% in predicting informational needs from survivor posts.
• The model effectively captures the expressed information needs within the text.
• This capability allows for rapid identification of support requirements.

Conclusions:

• Fine-tuned LLMs can accurately predict the informational needs of domestic violence survivors in online communities.
• This technology facilitates prompt and targeted support from healthcare providers.
• AI-driven analysis of online disclosures holds significant potential for victim assistance.