Abstract
During natural disasters, social media platforms serve as critical communication channels, yet Emergency Operation Centers (EOCs) often face severe information overload. This paper introduces Social-EOC, a framework designed to identify and prioritize "serviceable" requests—messages that are specific, actionable, and correctly addressed to responding agencies. By combining a Learning-to-Rank (LTR) framework with semantic grouping techniques, the system filters noise and clusters redundant information. Evaluations on six real-world crisis datasets demonstrate that Social-EOC improves ranking performance (nDCG) by up to 25%, significantly aiding emergency responders in identifying high-priority lifelines amidst digital noise.
Introduction
In the wake of a crisis, social media becomes a digital town square. While thousands of users post for help or share information, EOC personnel struggle to find actionable requests buried under a deluge of vague reports, expressions of gratitude, and rumors. The fundamental challenge is distinguishing between "relevant" content and "serviceable" content. A message like "Help is needed in Houston" is relevant but lacks the detail required for action. Conversely, a message specifying a location, a specific need (e.g., evacuation), and directed at a specific agency is highly serviceable.
The Social-EOC Framework
The researchers propose a multi-stage approach to automate the discovery of these high-value messages.
Defining Serviceability
Serviceability is defined through four key dimensions:
- Explicit Request (E): Does the message ask for a specific resource or service?
- Answerable Question (A): Is there a concrete question that can be addressed?
- Correctly Addressed (C): Is the message directed at an organization capable of providing help?
- Sufficiently Detailed (D): Does it provide necessary context (who, what, where, when)?
Ranking and Grouping
The system employs an SVM-Rank algorithm for prioritization. A unique feature introduced is "Sociability"—a proxy for user credibility based on the ratio of friends to followers. The hypothesis is that well-connected users are more likely to post credible, serviceable content to maintain their reputation.
To manage redundancy, the framework uses semantic grouping:
- Fine-grained clustering: Groups near-duplicate messages (e.g., retweets).
- Coarse-grained clustering: Uses the DBpedia ontology and word embeddings to bucket requests into functional categories like "Medicine" or "Transportation."
Experimental Setup
The model was validated using six crisis datasets (2012–2017), including Hurricane Sandy, the Nepal Earthquake, and the Louisiana Floods. A "gold standard" ground truth was established through collaboration with emergency professionals who labeled messages based on their operational serviceability.
Results and Discussion
The inclusion of serviceability dimensions and social features led to a significant performance boost, with Normalized Discounted Cumulative Gain (nDCG) improving by up to 25% over standard text-based baselines.
As shown in the correlation analysis, the proposed serviceability characteristics align strongly with expert judgments across diverse events. The model also demonstrated "cross-event robustness," performing well when trained on one disaster and tested on another—a critical requirement for real-time deployment in new, unlabeled crises. Furthermore, the semantic grouping approach effectively prioritized the most important clusters, achieving nDCG@5 scores of over 98%.
Conclusion
The Social-EOC model represents a shift from simple relevance filtering to operational serviceability. By quantifying the "work" required for a responder to act on a message, the system effectively reduces cognitive load. Future work will look to address language barriers, incorporate multi-level serviceability grading, and integrate real-time social network structures to further refine emergency response capabilities.
Summary updated