【摘要】 【目的】洪涝灾害是我国最严重的自然灾害之一，造成巨大的经济社会损失。【方法】面向洪涝灾害影响评估，本文构建基于大语言模型的新闻文本挖掘方法。首先，利用DeepSeek-R1-0528设计检索策略，通过慧科数据库获取洪涝灾害相关新闻，采用TF-IDF算法进行去重，形成新闻数据集；其次，构建涵盖18个维度的经济社会影响分类体系，经重复实验优选模型温度参数，提取新闻文本中的洪涝灾害影响信息；最后，结合官方发布的灾情统计数据验证提取结果，针对洪涝灾害案例展现灾情演变过程。【结果】从2024年的10 556篇新闻文本中提取出14 778条洪涝灾害影响信息，结果表明：大语言模型准确率和F1评分的中位数分别为0.91和0.73；洪涝灾害影响信息数量与省级行政区受灾人口数据之间的相关系数为0.68，其空间分布与2024年十大自然灾害中洪涝灾害的受灾地区基本吻合。并且，大语言模型有效捕捉到2024年4月粤北（清远、韶关）及6—7月初湖南岳阳暴雨洪涝灾情的动态演变，展现新闻焦点从灾中抢险到灾后恢复的转变。【结论】整体上，本文结果表明大语言模型可作为传统灾害调查评估的重要工具，对于灾情评估和应急决策具有重要的应用潜力与参考价值。更多还原

【Abstract】 [Objectives] Flood disasters are one of the most severe natural disasters in China, leading to huge socioeconomic losses. [Methods] This study establishes a flood disaster impacts assessment method based on news text mining via Large Language Models(LLMs). Firstly, retrieval strategies are designed using DeepSeekR1-0528 to obtain flood-related news from the Wise Search database, which are then processed using the TF-IDF algorithm for deduplication to form a news dataset. Secondly, a classification system covering 18 socioeconomic impact dimensions is constructed, and flood impact statements are extracted from news texts after optimizing model temperature parameters through repeated experiments. Finally, the extracted results are validated against official disaster statistics, and the dynamic evolution processes of disaster situations are presented based on specific flood cases. [Results] To estimate the socioeconomic impacts of the flood disasters in 2024, a total of 14 778 flood impact statements are extracted by the LLM from 10 556 flood-related news articles. The results show that the LLM performs well in the extraction task, with a median accuracy of 0.91 and a median F1 score of 0.73. At the provincial level, the number of flood impact statements is positively correlated with the official disaster data with a correlation coefficient of 0.68 and their spatial distribution is consistent with the areas of flood disasters in the top 10 natural disasters in 2024. Furthermore, the LLM effectively captures the dynamic evolution of the rainstorm and flood disasters in northern Guangdong(Qingyuan and Shaoguan) in April 2024 and in Yueyang, Hunan Province from June to early July, revealing a shift in news coverage from emergency responses to post-disaster recovery. [Conclusions] Overall, the results indicate that LLMs can serve as an effective tool for traditional disaster survey and evaluation, with important potential and reference value for postdisaster assessment and emergency decision-making.更多还原