EQUAL: Entity-Enhanced QUery Expansion for EquitAble Crisis Summarization via KnowLedge Graphs

Abstract

Disaster response efforts face persistent challenges in ensuring equitable aid and information access for all affected populations. Marginalized communities, including the elderly, persons with disabilities, people experiencing homelessness, low-income households, non-English speakers, and geographically isolated residents, face heightened risk during disasters and are more likely to experience delays in receiving aid, evacuation support, and critical information (Wilson et al. 2021). We present EQUAL (entity-enhanced query expansion for equitable crisis summarization via knowledge graphs), a work-in-progress framework that combines dual-model entity-enhanced query expansion with an equity-aware GraphRAG (Graph-based Retrieval-Augmented Generation) pipeline. EQUAL constructs crisis knowledge graphs enriched with vulnerability–resource connections and generates summaries through community-level synthesis. Evaluated on 18 real-world disaster events from the TREC CrisisFACTS dataset, EQUAL outperforms all baselines on equity-focused metrics—vulnerable group coverage, intersectional coverage, and statistical parity—and shows marked gains in explicit mentions of vulnerable populations, geographic specificity, and actionable resource information. It also remains competitive on standard semantic quality metrics.