LLM-guided Semi-Supervised Approaches for Social Media Crisis Data Classification

Researchers evaluate LLM-guided semi-supervised learning methods for classifying crisis-related social media data, finding that LG-CoTrain significantly outperforms traditional approaches in low-resource settings while compact models can rival large zero-shot LLMs. This demonstrates practical pathways for deploying AI in disaster response applications with minimal labeled training data.

This research addresses a critical gap in machine learning applications for crisis management. Social media platforms generate massive volumes of disaster-related content, yet labeling sufficient training data remains expensive and time-consuming. The study evaluates two LLM-assisted semi-supervised approaches—VerifyMatch and LLM-guided Co-Training—against classical baselines, establishing empirical benchmarks for practical deployment scenarios.

The findings reveal that LG-CoTrain achieves superior performance when working with extremely limited labeled data (5-25 examples per class), which mirrors real-world disaster response conditions where rapid deployment is critical. Notably, the research demonstrates that smaller, more efficient models trained using knowledge transfer from large language models can match or exceed the performance of large zero-shot LLMs—a significant finding for resource-constrained disaster management organizations that lack infrastructure for deploying billion-parameter models.

For the AI and crisis management sectors, this research validates a practical framework for knowledge distillation and efficient model deployment. Organizations responding to disasters need deployable solutions that work offline or with limited computational resources; this work demonstrates that LLM-guided semi-supervised learning bridges the gap between state-of-the-art performance and practical constraints. As labeled training data becomes available post-disaster, Self Training emerges as a strong alternative, enabling organizations to progressively improve models as human annotators categorize incoming data.

Future work should focus on cross-disaster generalization, multi-lingual crisis classification, and integration with emergency response systems. The GitHub repository release enables reproducible research and practical adoption by disaster management agencies.

• →LG-CoTrain outperforms classical semi-supervised methods in low-resource settings with minimal labeled training data per class.
• →Compact models trained via knowledge transfer from LLMs can match or exceed large zero-shot LLM performance while remaining deployable.
• →Semi-supervised learning approaches reduce the annotation burden for social media crisis classification in time-sensitive scenarios.
• →Performance gains from LLM-guided methods narrow as labeled training data increases, with Self Training becoming competitive.
• →Findings support practical disaster response applications requiring efficient, deployable AI solutions in resource-constrained environments.