Abstract: For the widespread and rapid information dissemination on social media, it is of great significance for automobile manufacturers to improve the product design and quality management. The current research on social media defect discovery can only mine a little part of valuable information and the approach is mainly clustering. Based on the existing research of defect discovery from social media, defect categories are extended according to the actual situation of the company; three feature selection methods are compared by Naïve Bayes classification method, and on the basis of this, a semi-supervised learning method using EM is applied. The experimental results show that the method proposed can effectively detect the vehicle defects to provide decision support for the defect management of the enterprise and reduce half the cost of manual tagging.

Key words: social media, defect discovery, feature selection, semi-supervised learning