Abstract: Catastrophes are happening more and more frequently, and the timeliness of social networks takes an increasingly important role in the rapid assessment of disaster situation and post-disaster recovery. This paper conducts a case study of the typhoon Lichma that happened on August 9, 2019 and collects 2.6 million relevant short-text messages posted from August 9 to 14, 2019 on the Sina Weibo platform. This paper compares the results of sentiment analysis on Weibo short texts using the machine learning and the optimized sentiment dictionary. Firstly, this paper finds that under the same time cost, the effect of machine learning for natural language processing is far inferior to the sentiment dictionary. The analysis accuracy of machine learning is greatly affected by both the topic and the corpus. Then, this paper uses the sentiment dictionary based on typhoon to analyze the sentiment of the short-text of Weibo, and finds that the sentiment curves of the severely damaged provinces show greater fluctuations, and the emotional fluctuations caused by personal injuries are much greater than those caused by property losses. In the Linhai event, there was a two-hour time lag between the ‘good’ emotional curve and the ‘fear’ emotional curve, implying that the peak of the ‘rejoicing’ mood is a sign that there could be more serious damage to come soon, and the time lag provides a new idea for disaster early warning.

Key words: social media, natural language processing, typhoon Lekima, disaster warning