Due to the Internet and mobile terminal technology’s rapid development, significant changes have taken place in culture, politics and social interaction on network. The rapid emergence of new social platforms like Weibo, Blogs and WeChat has steadily altered people’s perceptions on information, the lower bar in informational distribution makes it easier for rumors to diffuse. At present, the majority methods employed in rumor detection focused on classifications and regional traits extraction, while ignored the emotional signals in publishers and receivers. So, we proposed an improved model based on dual sentiment features and commentators, then embedded in CNN, RNN and BERT network for rumor detection. These given results have demonstrated that the methods mentioned above are higher than 80% in terms of accuracy and F1. Besides, it also reflects a higher accuracy and better detection effect, compared with the single semantic detection model.