基于社交媒体与大模型的城市内涝灾情信息提取研究OA

Urban underlying surfaces are complex and infrastructure is inadequate,making cities prone to waterlogging disasters during heavy rainfall.Rapid identification and extraction of disaster information is crucial for improving urban flood response efficiency and disaster management capabilities.Social media,due to its real-time nature and high public participation,has become an important data source for obtaining urban waterlogging disaster information.To address challenges such as the large scale,semantic complexity,and difficulty in directly supporting decision-making of social media disaster text information,this study takes the Zhengzhou"7·20"rainstorm as an example and constructs a 15-category fine-grained classification system based on the urban waterlogging disaster chain.A large language model(LLM)-driven disaster information extraction method is proposed,with the BERT-TextCNN model introduced as a benchmark for comparative analysis.The results show that:①LLMs exhibit high few-shot learning efficiency.The LLMs DeepSeek,ChatGPT,and GLM can achieve disaster classification with small sample sizes,and their performance shows a positive correlation with data scale.Among them,ChatGPT performs best,achieving a macro-and micro-averaged F1 score of up to 0.89.②The deep learning benchmark model performs robustly with larger sample sizes.BERT-TextCNN demonstrates reliable classification capability with sufficient labeled data,achieving a weighted F1 score of 0.86.③LLMs have a significant data efficiency advantage.With only 450 labeled samples,DeepSeek and ChatGPT achieve comprehensive performance superior to the BERT-TextCNN model trained on 5 600 samples,significantly reducing the cost of manual annotation.This study can provide a reference for urban waterlogging disaster information extraction and emergency decision support.