Knowledge Driven Event Extraction from Low Resource Language Content Using Strategic Few-Shot Learning

Timely extracting structured information from social media during disaster events is critical for effective response and coordination. This work presents a comprehensive pipeline for sub-event information extraction (IE) from Hindi-language tweets and news articles, covering eleven key categories such as casualties, infrastructure damage, relief operations and others. We introduce CrisisIE-Hindi, a benchmark dataset of around 9.6k annotated posts spanning cyclone, earthquake, flood and other disaster events. To assess the robustness of disaster IE models under low-resource and multilingual settings, we evaluate seven open-source and proprietary language models of varying sizes. Our analysis reveals that smaller models struggle to follow structured output instructions and generalize across unseen scenarios, particularly those with limited exposure to Hindi or domain-specific data. To mitigate these challenges, we propose a strategic few-shot prompting framework that leverages a frozen data bank for dynamic demonstration selection. This approach significantly improves model performance across all eleven categories, even on out-of-domain events. To show the efficacy of our proposed approach, we also showed the performance on three other low-resource language datasets- L3-Cube-MahaNER, B-NER, and HiNER. Notably, our method enables small and mid-sized language models to generalize effectively in instruction-following tasks for disaster IE. Our findings highlight the potential of prompt-based adaptation for low-resource languages and domains. We release our dataset to foster future Hindi disaster response systems research.