As an important research area of natural language processing, text sentiment analysis has been widely used in opinion processing and product review analysis. Recurrent neural networks have the characteristics of temporal order, cannot directly extract the contextual semantic features of sentences, and cannot be computed in parallel; convolutional neural networks can be computed in parallel, but have the disadvantages of huge computation and long training time. To address the above problems, this paper proposes a Light-Transformer model based on word vector representation and positional embedding, and modifies the structure of the Transformer to keep only the Encoder module, and finally the extracted sentence features are input to two fully connected layers for classification. Experimental results on the NLPCC2014 Task2 dataset show that the method improves the classification accuracy by 0.3%-1.0% compared to traditional methods such as LSTM, CNN, etc., and the number of model parameters is greatly reduced compared to other Transformer-based models with close accuracy.