AAI_2025_Capstone_Chronicles_Combined

2

food, or shelter, which requires careful categorization to ensure appropriate resource allocation and timeline intervention. However, the manual triage of these high-volume, time sensitive communications creates a significant bottleneck that can delay life-saving responses. This challenge has been further compounded by recent funding reductions to federal weather and natural disaster response agencies, placing additional strain on already overwhelmed emergency management systems (Ndugga et al., 2025). As response organizations face growing demands with fewer resources, the need for automated, scalable solutions to process disaster communications has become increasingly critical. Manual categorization of emergency messages is not only time-intensive but also prone to inconsistency and human error, particularly under the high-stress conditions that characterize disaster response operations (Castillo, 2016). This project addresses the urgent need for automated disaster message classification by developing a multi-label machine learning system capable of categorizing emergency communications across 36 distinct action categories. Unlike traditional single-label classification problems, disaster messages frequently contain multiple simultaneous needs (a single message might indicate needs for medical assistance, water, and shelter) necessitating a multi-label approach that can identify all relevant categories for each communication (Burel et al., 2017). By leveraging natural language processing (NLP) techniques combined with multi-label classification algorithms, this system aims to enable near-instantaneous message triaging, allowing response organizations to route communications to appropriate teams within seconds rather than hours.

The primary end users of this AI system include disaster response organizations such as the American Red Cross, The Federal Emergency Management Agency (FEMA), non-governmental humanitarian organizations, emergency operations centers, and first responders coordinating relief efforts. Additionally, local communities and grassroots organizations managing disaster response at the municipal level would benefit from this automated categorization capability. These stakeholders require immediate, accurate categorization of incoming messages to prioritize the most urgent needs and to allocate limited resources effectively across geographic areas and need types. This project utilizes the Figure Eight Disaster Response Messages dataset, a comprehensive collection of approximately 30,000 real emergency messages gathered from major disaster events including the 2010 Haiti earthquake, the 2010 Chile earthquake, the 2010 Pakistan floods, and Hurricane Sandy in 2012 (Appen, 2020). These messages have been manually annotated across 36 binary categories representing different types of disaster-related needs and infrastructure concerns. The messages originate from diverse sources including direct communications, social media platforms, and news reports, providing a realistic representation of the heterogeneous communication channels emergency responders must monitor during actual disaster scenarios. Each message is labeled for multiple simultaneous categories, reflecting the complex, multi-faceted nature of disaster response needs. In a fully deployed production system, this model would integrate with real-time data streams from social media APIs, emergency SMS gateways, disaster response hotlines, and

318