With the rapid development of technology and online social networking, the popularization of smartphones has promoted the research and development of sentiment analysis of contemporary literary texts. In this paper, the CBOW model based on Hierarchical Softmax algorithm is used to extract text sentiment features. The classification mechanism of sentiment lexicon, machine learning, and deep learning methods supported by sentiment features is discussed. According to the discussion results, a 5-layer sentiment analysis model based on CNN-BiLSTM-ATT is built based on text preprocessing, and the model design of different layering is proposed. Meanwhile, the analysis method of text themes is proposed based on LDA. In the long story dataset, the model recall rate of this paper is 83.91% and the precision rate is 83.86%, the values are higher than the other six models; the MacroF1 mean value is 83.16%, which proves that the fused and improved CNN-BiLSTM-ATT model of this paper possesses excellent performance in the sentiment analysis task. In the short story dataset, the accuracy, precision and recall are not less than 98%, and the loss rate is the lowest 34.11%, which are lower than the other six models. The model in this paper can be applied to text analysis systems and has superiority in parsing the sentiment of contemporary literature.