Estimation of Danger Signs in Regional Complaint Data

Abstract

Government 2.0 activities have become very attractive and popular. Using the platforms to support the activities, anyone can anytime report issues in a city on the Web and share the reports with other people. Since a variety of reports are posted, officials in the city management section have to give priorities to the reports. However, it is not easy task for the officials to judge the importance of the reports because importance judgments vary depending on the officials, and consequently the agreement rate becomes low. To remedy the low agreement rate problem of human judgment, it is necessary to create an intelligent agent which supports finding reports with high priorities. Hirokawa et al. employed the Support Vector Machine (SVM) with a word Feature Selection method (SVM+FS) to detect signs of danger from posted reports because the signs of danger is one of high priority issues to be dealt with. However they did not compare the SVM+FS method with other conventional machine learning methods and it is not clear if the SVM+FS method has better performance than the other methods. This paper explores methods for detecting the signs of danger through comprehensive experiments to develop an intelligent agent which supports officials in the city management sections. We explores conventional machine learning methods: SVM, Random Forest, Naïve Bayse using conventional word vectors, an LDA-based document vector, and word embedding by Word2Vec and compared the best method with SVM+FS. Experimental results illustrate the superiority of SVM+FS and invoke the importance of using multiple data sets when evaluating the methods of detecting signs of danger.