Data-centric AI in disaster management: Scalable Named Entity Recognition via LLM-based stacking and pseudo-label generation

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).

This study presents a scalable Teacher–Student architecture for Named Entity Recognition (NER) that enables automatic information extraction from Turkish disaster-related social media texts. To reduce labeling costs, the collective intelligence of three large language models (Gemini 2.5 Pro, Claude Sonnet 4.5, GPT-5) is combined through a learnable stacking meta-classifier. On a 500-tweet manually annotated dataset, stacking reached a token-level Macro F1 of 0.9313 (0.9219 under leakage-free tweet-level grouped cross-validation), on par with the strongest individual model (0.9245) and clearly above a BERT baseline trained on the same data (0.8335). It was selected as the Teacher for its learnable per-class weighting, its stability across meta-classifiers and its robustness to single-provider outages. The Teacher’s discrete predictions were transferred to a more efficient BERT Student through hard-label pseudo-labeling, and this Student was then used as a zero-cost proxy labeler to scale the corpus to 17,000 tweets. On a fully isolated 300-tweet blind test set, the final model obtained an entity-level Macro F1 of 0.9160, against 0.8648 for the baseline trained only on manual labels. A size-matched controlled experiment isolates the source of this gain: labeling the same 3000 tweets with the model ensemble instead of a self-training labeler improves entity-level Macro F1 by 0.033 in all five random seeds ( p < 0.001 ), whereas scaling from 3000 to 17,000 instances adds only 0.006. Label quality, rather than data volume, drives the improvement, and the approach offers a practical, low-cost route to rapid humanitarian needs mapping when annotated data is scarce.