Multiclass Classification and Spatial Clustering of Disaster Tweets for Humanitarian Awareness

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Background: Social media is a critical channel of real-time information channel during natural disasters, but the informal and fragmented nature of tweets makes automatic classification difficult. Named entity recognition (NER) can capture contextual cues, such as affected locations. However, how entity information should be incorporated into classifiers and its value for disaster mapping remain underexplored. Objective: This study investigates whether NER can improve disaster-tweet classification and support clustering-based disaster mapping for humanitarian decision support. Methods: Eight configurations were evaluated using raw text: linear and BiLSTM classification heads combined with three input representations (Raw, Inline Tag, and Special Token), and two architecture-level injection models (BiLSTM+TagEmb and Two-Stream) using raw text. An ablation study compared eight pretrained language models across four architectures and input variants. In addition, six density-based clustering algorithms were applied to NER-extracted location entities to map the spatial distribution of disaster events. Results: The two-stream model achieved the highest performance (macro F1 = 0.947), followed by the BiLSTM classifier with raw input (0.914), indicating that the architecture-level entity integration outperforms the input-level entity marking. VDBSCAN, HDBSCAN, and OPTICS produced the most cohesive clusters (silhouette > 0.97). Conclusion: Integrating entity information at the architectural level is more effective than embedding entity markers in the input text. Density-based clustering is better suited to the exploratory discovery of spatial patterns than to reproducing predefined categories. These findings provide practical guidance for designing disaster-tweet classifiers and selecting unsupervised methods to analyze disaster-related social media data. Although the proposed approach may contribute to situational awareness, its utility for operational humanitarian decision-making remains to be empirically validated. Keywords: Disaster tweet classification, Named entity recognition, natural language processing, social media analytics, disaster