# A collaborative taxonomy of social media indicators for localised disaster response

## TL;DR

This paper introduces a context-specific taxonomy of social media indicators for disaster response, developed through stakeholder collaboration in Ghana and Mauritius.

## Contribution

The study proposes a participatory framework for creating adaptive, region-specific social media taxonomies to improve disaster response AI systems.

## Key findings

- A taxonomy of 39 social media indicators was developed across four categories for localized disaster response.
- Regional variations in disaster information priorities contradict universal AI assumptions.
- Participatory methods offer a replicable approach to align AI systems with local needs.

## Abstract

Effective disaster management hinges on prompt, informed decisions, where social media has emerged as a real-time information source. However, current artificial intelligence (AI) systems for disaster response rely on universal taxonomies that assume information relevance is consistent across geographical and cultural contexts – an assumption that fails to account for regional variations in disaster types, response capabilities and local priorities. This study questions the ‘one-size-fits-all’ approach by developing context-specific social media indicator taxonomies through participatory engagement with 104 stakeholders across Ghana and Mauritius. We developed a taxonomy of 39 social media indicators across four categories: urgent needs, impact assessment, situational awareness and vulnerable populations. Our findings reveal significant regional variations in disaster information priorities that contradict assumptions underlying existing universal frameworks. While impact assessment indicators showed convergence between countries, other categories revealed that there are still important areas for future research on incorporating local stakeholder knowledge into AI system design. Our participatory methodology provides a replicable framework for developing adaptive, context-aware machine learning classifiers that can transform static universal categorisations into dynamic systems aligned with unique regional priorities and operational contexts.

We suggest future research areas that span across developing transfer learning approaches that leverage pre-trained multilingual models while incorporating region-specific context, creating active learning frameworks with local validation loops, implementing feedback mechanisms and establishing fair human-in-the-loop annotation processes that maintain quality.