Improving Emotional Support Delivery in Text-Based Community Safety Reporting Using Large Language Models
Project Overview
The document explores the application of generative AI, particularly Large Language Models (LLMs), in enhancing emotional support within text-based community safety reporting systems at universities. An analysis of two years of chat logs revealed that human dispatchers often experience a decline in the consistency and quality of emotional support over time. In response, the introduction of dispatcherLLM—a fine-tuned LLM—aims to provide more reliable and effective emotional support during incident reporting. The findings indicate that while human interactions may fluctuate, dispatcherLLM consistently offers improved emotional support, thereby enhancing service delivery in community safety contexts. The document illustrates various dispatch scenarios where users report safety concerns, showcasing how AI can facilitate communication, bolster emotional support, and streamline response efficiency. Additionally, it presents examples of user interactions with AI dispatchers, highlighting the potential benefits of integrating such technology into community safety systems, as well as the challenges that may arise during implementation. Overall, the integration of generative AI in education-related safety reporting systems not only aims to improve emotional support and communication but also seeks to transform the efficiency of response mechanisms.
Key Applications
dispatcherLLM (Large Language Model) for Community Safety Reporting
Context: Text-based community safety reporting systems in higher education institutions and local police departments, where individuals report incidents and require emotional support during these interactions.
Implementation: Fine-tuned AI models, specifically large language models, are utilized to generate human-like responses and provide emotional support in safety reporting contexts. These models are trained on chat logs from applications like LiveSafe and similar platforms to enhance their effectiveness.
Outcomes: Improved communication between users and dispatchers, enhanced emotional support for individuals reporting incidents, faster response times, and overall consistency and effectiveness in emotional support compared to human dispatchers and other LLMs.
Challenges: AI may struggle with understanding context, generating appropriate emotional responses, ensuring privacy and data security, and may lack domain-specific emotional expertise in sensitive contexts like mental health. Additionally, human dispatchers may show a decline in delivering emotional support over time.
Implementation Barriers
Organizational
Decline in emotional support provided by dispatchers over time as organizations gain experience with the reporting system. Higher volumes of incident reports during certain shifts lead to dispatcher burnout and reduced emotional support.
Proposed Solutions: Regular assessments and training for dispatchers to maintain or increase emotional support. Assign dedicated categories of incidents to dispatchers based on complexity and emotional support needs.
Technical Barrier
AI may not fully comprehend complex human emotions or context in emergency situations, leading to inadequate support.
Proposed Solutions: Continuous training and fine-tuning of AI models, incorporating feedback from human dispatchers to improve AI responses.
Privacy Barrier
Concerns regarding the handling and storage of sensitive user data during interactions.
Proposed Solutions: Implement strict data privacy policies, anonymize user data, and ensure compliance with legal standards.
Project Team
Yiren Liu
Researcher
Yerong Li
Researcher
Ryan Mayfield
Researcher
Yun Huang
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Yiren Liu, Yerong Li, Ryan Mayfield, Yun Huang
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai