Towards Location Approximation of Typhoon Related Discourse: On Region Definition and Temporal Segmentation
Tweets have augmented disaster information in variety of ways. Different tools employ tweets as a data source to aid in disaster event detection, knowledge extraction and situational awareness. A constant problem faced by these efforts, however, is the lack of geospatial information in majority of tweets leaving only less than 1% of tweets useful for data mining. Though studies have devised methods to approximate the location of non-geotagged tweets, applying these same methods on disaster tweets may not be the optimal choice given that disaster tweets contain domain specific characteristics. A model for approximating the location of disaster tweets must take into account how human discourse changes as the typhoon progresses, and therefore, how tweet content is affected by the location of the eye of the typhoon and the disaster affected areas. This study seeks to find the geospatial characteristics of disaster tweets in relation to typhoon relevant locations and to present initial models for predicting disaster related tweet locations through region definition relative to disaster relevant locations such as the disaster affected area and the eye of the typhoon. The first explores characteristics and relationships between the path of the typhoon, the location of the affected area and the location of the tweets pertaining to the typhoon. The second part presents two new disaster relative region definition schemes, namely: regions defined based on relative distance to the disaster affected area and regions defined based on relative distance to the eye of the typhoon. Tweets can then be geotagged to these defined regions. Based on the results, the location of the disaster related tweets are significantly related to both the location of the typhoon and an identified affected area. Furthermore, models that predict location relative to the disaster area outperform models that predict location relative to the eye of the typhoon both in accuracy and error distance. The results also show that for geospatial modeling, there is a need to consider creating models for each specific and smaller timespan instead of a single model for the whole coorpus to increase accuracy and lower error distance.
John Clifford Rosales and Ma. Regina Estuar. 2017. Towards Location Approximation of Typhoon Related Discourse: On Region Definition and Temporal Segmentation. In Proceedings of the 9th International Conference on Machine Learning and Computing (ICMLC 2017). Association for Computing Machinery, New York, NY, USA, 482–492. DOI:https://doi.org/10.1145/3055635.3056624