Tone and sentiment in financial contexts，containing emotional information expressed by managers in public listed firms and individual investors，affect stock market． By restructuring dictionaries and using deep learning model LSTM，the paper constructs two Chinese sentiment dictionaries for formal and informal texts in Finance respectively． Based on the constructed dictionaries，tone measures of annual filings and sentiment proxies of social media for Chinese public firms are proposed． Our tone measures of annual filings and sentiment proxies of social media can effectively predict stock return，trading volume，return volatility，unexpected earnings and other market factors and perform better than indices made by other commonly used sentiment lexicons． Additionally，our tone measures of annual filings and sentiment proxies of social media have predictive abilities for crash risk of public firms． This research provides an analytical tool for big data application in financial market and offers decision-making supports in financial market forecasting，monitoring，and other activities in the big data era．