This study explores the effects of textual tone of analyst reports on the stock market information efficiency and its mechanisms. On the basis of 326,425 sell-side analyst reports of Chinese listed firms from 2006 to 2017, 10,434 clauses from analyst reports are randomly selected and these training materials are manually classified these into positive, neutral or negative category. Then 11 machine learning methods including Na?ve Bayes, logistic regression, neural network, etc., are applied to predict the tone of each clause based on the training materials. Predictive accuracy of 11 machine-learning methods is compared. Na?ve Bayes, which proves to be the best method of forecast, is finally used to measure the analyst report tone. The empirical results show that the tone of analyst report is negatively associated with the stock price synchronicity of the companies. Although the results are different from most of conclusions of the existing researches, they can be well explained by the individual selective perception theory in the capital market of China, where the short-selling mechanism is underdeveloped. Furthermore, we analyze three mechanisms of the above negative effects and test the mediating effects of stakeholder behaviors. The results indicate the analyst’s positive textual tone improves the information efficiency of the stock market by (1) stimulating the firms to issue more announcements, (2) guiding institutional investors to buy and (3) attracting other analysts to release more reports. The above conclusions remain tenable in multiple robustness tests. This study has important implications for investors to focus on the index of textual tone, for listed companies to strengthen the information disclosure management, and for government departments to improve the capital market system.