Because of the convenience of the Internet, there are many websites or online news spread misinformation, cause panic and trepidation in society. Automatic fake news detection can classify fake news and help the society to clarify the information is true or false without human checking. Detecting fake news by analyzing the stance is one of the mainstream methods, stance detection has become a new popular research field in recent years. How to accurately detect stance has become the primary goal of detecting fake news. This research aims to detect the news stance accurately, and we propose a method based on a pre-trained BERT language model. Most of the previous work only used the knowledge of single inference direction when classifying the stance, which may lose some important information. Therefore, we propose a bidirectional inference stance detection model, which can leverage bidirectional perspective information to classify the stance more comprehensively. We also define the stance detection task as a hierarchical structure task, and use the hierarchical classification and incorporate the topic information to help the stance classification. Experiment results show that our model can classify the stance more accurately.