چكيده به لاتين
In recent years, Disaster management is one of the topics that has always attracted the attention of researchers, especially civil engineers. On the other hand, today, the context of social media and the advent of artificial intelligence has created countless opportunities to gather instant information and examine society's response to crises, so that the general public has important information such as early warnings, requests for assistance and shortages, service satisfaction, infrastructure damage such as roads, bridges and buildings are reported in this context. Some challenges, such as content diversity, high production speed and volume, and the presence of irrelevant content, make it impossible to use previous methods to process and use this data. Therefore, in this dissertation, we focus on providing a framework for analyzing Twitter data in the event of a disaster, which includes the use of modern natural language processing and machine learning methods to analyze and process this data with the aim of creating a disaster management tool. In the framework presented with the knowledge extraction approach, the received Twitter data, which includes specifications and metadata, is first stored in the database and then enters the data mining stages in the analysis of text, network, place and time, and finally the results of these analyzes are combined. For topic and sentiment classification, we use transformer-based text processing models and transfer learning approach, which is one of the state-of-the-art approaches in this field. The trained models for analyzing positive and negative emotions, topic classification for 2 classes and classification of 11 classes, showed accuracy of approximately 91, 87 and 86% in the test, respectively, which was found to be very accurate in the field of crisis informatics. After analyzing the Dorian hurricane data as a case study, the relevant results and interpretations were presented, some of which are: tweets with the subject of warning and advice before the accident grew rapidly and decreased from the moment of the accident onwards, and tweets with the content of the request at the day of the accident suddenly intensified and continued the next day, but the tweets about the damage to the infrastructure reached higher levels almost a day after the accident, which is a sign of the delay in the community's response to this issue. Also, most of the cities where the wave of tweets suddenly intensified were on the hurricane route. The shorter the distance from the incident at the time of the incident, the more tweets were published from that location. In summary, the most important application of the proposed model is access to software to process existing information to help achieve rapid insight into the current situation for prioritization and decision-making or performance appraisal, public satisfaction and recognition of effective factors and trends for appropriate response to future or ongoing crises.
