Enhancing Disaster Response through Named Entity Recognition of Critical Infrastructure and Medical Resources

Yuchen Tang^*

• Chongqing University of Science and Technology, Chongqing, China

The increasing frequency of disasters presents significant challenges for emergency response systems, particularly in identifying and allocating infrastructure and healthcare resources from unstructured data. Named Entity Recognition (NER) offers a solution but struggles with noisy inputs, sparse labels, and temporal variability. This paper proposes a novel framework combining the Tectonic Entity Recognition Network (TERN) and the Seismic Refinement Mechanism (SRM). TERN utilizes a multi-scale transformer with segmentation-aware attention and temporally conditioned decoding to enhance semantic coherence in dynamic contexts. SRM refines predictions through soft re-scoring using lexical, semantic, and temporal signals, improving robustness to rare and evolving entities. The framework integrates consistency learning and cross-document alignment to maintain performance across variable data streams. Experiments on disaster-related datasets show notable gains in recall and F1 scores, especially for infrastructure and medical entities, demonstrating the framework's potential to strengthen situational awareness and resource coordination in high-volatility scenarios.