Estimating Priority of Chronology Logs in Earthquakes and Torrential Rains

The Disaster Response Headquarters receives various requests and reports from affiliated organizations, such as hospitals.Staffs at the headquarters need to prioritize incoming information to effectively allocate limited resources, including supplies and personnel.To organize information chronologically and prioritize responses appropriately, staffs use format called chronology logs to avoid confusion.However, given the large volume of information received, prioritizing entries in the chronology logs is a demanding task.To alleviate this burden, we aim to develop an automatic estimation system for the priority of chronology logs using a deep neural network model.In this study, we classify the priorities in the chronology logs into three levels: high, medium, and low.A dataset annotated manually with these priority levels was created using actual chronology logs to train the model.The chronology logs dataset is imbalanced, with approximately 90% of samples in the low-priority class.In this paper, we compare methods such as weighting, data replication, and data augmentation in training a BERT model on the chronology logs dataset to identify effective techniques to mitigate the adverse effects of class imbalance.We also compare priority estimation models using BERT and GPT embeddings to verify which is most effective for estimating priority of chronology logs.Experimental results show that data replication