Robust Zero-Shot NER for Crises via Iterative Knowledge Distillation and Confidence-Gated Induction

Keywords: Zero-Shot NER, Crisis NLP, Iterative Knowledge Distillation, Confidence Gating, Gazetteer Induction, Rule Induction, PLMs

Abstract: This research explores the brittleness of Named Entity Recognition (NER) in cold-start crisis scenarios, where models often fail to adapt to novel disaster lexicons without manually curated resources or task-specific supervision. A confidence-gated iterative induction framework is introduced to address this challenge. It leverages a pretrained language model to extract high-recall entity candidates, then iteratively distills domain knowledge through a self-correcting loop that uses high-confidence seeds to induce micro-gazetteers and syntactic rules. These resources refine and update entity predictions. Evaluations on data simulating crises through leave-one-event-out protocols reveal that the framework maintains a constant zero-shot F1-score of roughly 0.295 with current hyperparameter settings, indicating that the iterative mechanism provides no measurable improvement in its current form. Nevertheless, this approach offers interpretable knowledge for disaster response and highlights practical limitations, such as error propagation risks and the difficulty of adapting to unreliable early seeds. The findings affirm the complexities of achieving robust zero-shot NER in real-world crises and underscore the need for future refinements.

Submission Number: 115