<p>This study constructs a multi-level investor sentiment index through the integration of multi-source heterogeneous data, thereby significantly enhancing the accessibility of sentiment data. Spanning the sample period from January 7, 2010, to October 30, 2020, the research employs a composite index method to measure investor sentiment using traditional indicators such as turnover rate, price-earnings ratio, and the advance-decline line. Furthermore, natural language processing techniques are utilized for sentiment classification to extract emotional tendencies from online posts, thereby constructing a more comprehensive investor sentiment index. The study delves into the intricate relationship between investor sentiment and stock market returns from both time-frequency and asymmetric perspectives. On this basis, a DBN-BP prediction model is developed to examine the effectiveness of the investor sentiment index in forecasting stock market returns. The results show that there is a positive and negative interaction between the comprehensive sentiment index and stock market returns on the medium to long term scale, while the text mining sentiment index and stock market returns show a strong positive correlation on the short and medium term scales, and there is an asymmetric cross correlation between investor sentiment and stock market returns. Additionally, the DBN-BP neural network model proposed in this study demonstrates superior prediction performance compared to the BP neural network and RBM model. This also confirms the predictive utility of the investor sentiment index constructed from multi-source heterogeneous data fusion for stock market returns.</p>