The research of stance detection aims to study the support, neutrality or opposition expressed by specific texts on specifictopics. In the previous research?on the stance detection of Chinese texts,the dependence between text structures was not paid attention to,and the stance implied in the comment text was often obscure and insensitive. We propose a stance detection method combiningBidirectional Encoder Representation from Transformers,Long Short Time Memory ( LSTM) and Convolutional Neural Network ( CNN)to solve this problem. We propose an optimal word number dimension determination algorithm to analyze the input of BERT model. Onthe model construction,the parallel input and output method is used innovatively, making full use of the advantages of LSTM globalfeature extraction and CNN local feature extraction. The BERT model can extract more obscure features and insensitive features. Thismethod can effectively determine the support,neutrality or opposition expressed by different targets for a specific topic. Compared withtraditional models and existing stance detection methods,the proposed model has excellent performance,and its F1 value reaches 88. 3% .