Analysis of Disaster Scene Narratives Generated by Language Models to Provide Relevance with Individuals

Abstract

Despite the fact that Japan is a country prone to natural disasters, the participation rate in evacuation drills, especially among young people, has remained flat in recent years. This is considered the fact that it is difficult to accept disasters that have occurred in other regions as their own. As a result of preliminary experiments, it became clear that viewers tended to be less likely to feel it as their own if they simply watched disaster videos.

The purpose of this study is to generate disaster site episodes that viewers feel as their own, and to clarify the characteristics of the episodes that are common to them. To avoid black boxing, we used a small AttentionSeq2Seq model. We selected tsunami and landslide disaster site episodes that viewers felt as their own, generated using ChatGPT based on a questionnaire, and trained the model to output sentences that were identical to the input sentences. After training, the test data was fed into the trained model and we analyzed the generated episodes and the attention texts based on the attention maps and examined the final layer of the encoder. As a result of the analysis, it was confirmed that the model captures the factors that viewers feel as their own for each disaster type. By including these elements in the training data, it was shown that even a small amount of data could generate disaster episodes that felt more familiar to the viewer.