Automatic Identification of Helpful Information on Social Media During Natural Disaster Based on Word2Vec and Bert
Lei Shi, Daiying Zhao · 2023
Social media plays an important role in natural disaster rescue, which facilitates the government's understanding of situation awareness. In the event of disaster, the amount of assistance-related information published on the network has increased significantly, and the workload of manual screening and identification is huge. Automatic identification of help-seeking and help-providing information is crucial for effective response to disaster relief work, which can greatly reduce the burden of human resources and enable more human resources to be used for rescue work. And the all-weather work of the model can ensure the timeliness of rescue and greatly improve the speed of emergency response. However, data overload and the full of irrelevant, misleading, or inaccurate information in social media make it challenging to extract meaningful insights. In this paper, we collect the dataset of 2021 Henan rainstorm under the rescue-related hashtags, including 38,007 microblogs from Weibo, a Twitter-like social media platform in China. Firstly, we combine text explicit features (bag-of-words) with domain-expert knowledge to construct a labeled dataset. Then, based on Word2Vec and Bert, word embedding techniques are used to explore the implicit semantic features of microblogs. Then, we employ SVM, TextCNN, LSTM, BiLSTM and Transformer with different text feature to conduct comprehensive comparison. Finally, the optimal model is determined as BiLSTM-bert with the highest accuracy 98.87%. This study has substantial academic and practical value in improving government disaster rescue efficiency.