Extracting Useful Information for Disaster Response by Paraphrasing Tweets Using Sequence-to-Sequence Models

Abstract

The collection and dissemination of information are important to limit damage in times of disaster, and the use of social media is being promoted. However, the volume of information circulated during a disaster is enormous, and it is necessary to screen posts mechanically. The selection process involves aspects such as whether a particular topic, such as a disaster situation or a request for rescue, can be identified and whether the point of reference can be identified. We are trying to solve this problem as a task of document classification and unique expression extraction to the contributions, and it is possible to sort the contributions with a system that connects several models. On the other hand, it has the problem of increasing system complexity. In other words, the problem is complicated by the fact that the mechanical processing of natural sentences differs for each perspective. In response, the transformation of natural sentences into machine-processable sentences is verified. By transforming disaster tweets into a machine-readable form using a series transformation model, we attempt to extract the information necessary to select posts that are useful for disaster response.