The purpose of this study is for the problem that the existing language knowledge and emotion resources are not fully utilized in the emotion analysis tasks, as well as the problems in the sequence model:the model will decode the input text sequence into a specific length vector, if the length of the vector is set too short, the information of input text will be lost. A bidirectional LSTM sentiment classification method is proposed based on multi-channel features and self-attention (MFSA-BiLSTM). This method models the existing linguistic knowledge and sentiment resources in sentiment analysis tasks to form different feature channels, and uses self-attention mechanism to focus on sentiment information. MFSA-BiLSTM model can fully explore the relationship between sentiment target words and sentiment polar words in a sentence, and does not rely on a manually compiled sentiment lexicon. In addition, this study proposes the MFSA- BiLSTM-D model based on the MFSA-BiLSTM model for document-level text classification tasks. The model first obtains all sentence expressions of the document through training, and then gets the entire document representation. Finally, experimental verifications are conducted on five sentiment classification datasets. The results show that MFSA-BiLSTM and MFSA-BiLSTM-D are superior to other state-of-the-art text classification methods in terms of classification accuracy in most cases.