Employing a novel text mining method proposed by Ke et al. (2019), I carry out a comprehensive textual analysis of Chinese company announcements for stock return prediction. The long-short portfolio based on article sentiment scores yields 78% cumulative returns with a Sharpe ratio of 0.86 from 2016 to 2019. The model well captures the contemporaneous association between announcement sentiment and stock returns. There is also evidence for the information leakage before the announcements are made public. Careful examination of the traded articles finds that they are longer than average, suggesting complex information may not be fully digested by the investors as they are released. Company annual reports and financing activities reports play important roles in explaining the portfolio's performance.