Defect reports play an important role in software maintenance. Software engineers often process severe defect reports in their first priority. Therefore, severity prediction of defect reports becomes an important research issue. Although past studies have discussed the representative indicators of the severity, they do not discuss the usage of the indicators to improve the prediction performance. This research addresses this issue with six feature selection methods to extract both severe indicators and non-severe indicators. The experimental results show that appropriately reweighting these indicators can improve the prediction performance.