Keymines: Extracting Minimal Keyphrases for Sub-Events in Disaster Situations

The substantial volume of unstructured social media data generated during disasters often conceals critical information. Developing efficient methods to extract actionable insights from this data can significantly enhance emergency response and resource allocation. However, existing methods, primarily reliant on supervised learning, encounter challenges such as dependence on labeled data, limited adaptability, and scalability. To overcome these limitations, we present KeyMinES, an unsupervised model that extracts minimal keyphrases—bigrams and tokens—from social media data to identify and classify critical sub-events. Our approach integrates semantic and grammar-based reconstruction to ensure that the extracted keyphrases are both grammatically correct and contextually meaningful. Through clustering, we group these reconstructed sub-events, enabling the identification of patterns and offering actionable insights for decision-makers. Our experimental results, attained through quantitative and qualitative evaluations, demonstrate that KeyMinES outperforms baseline methods, achieving higher F1 scores and providing a scalable and cost-effective solution. Our ablation study reveals that combining bigram+token enhances sub-event detection compared to using only bigram or token, capturing both contextual relationships and granular details, thereby leading to more accurate identification of critical sub-events. This model holds significant potential for various stakeholders, including emergency responders and humanitarian organizations, by improving the extraction of actionable insights during disasters.