Abstract: Selecting the sample of Shanghai and Shenzhen A-share listed companies from 2006 to 2016, this paper examines whether listed companies would manipulate text information disclosure and use more subjective textual description to cover up their earnings manipulation behaviors. Machine learning is used to extract 2-POS subjective model and construct quantifiable subjective indicators. Empirical results show that before financial restatement, that is, when earnings manipulation is concealed, listed companies use more subjective sentences in their annual reports and have higher scores with regard to overall subjectivity;developed external market environment and strong corporate governance effectively prevent listed companies taking advantage of subjective text information to work for earnings manipulation. This paper provides a new evidence for motives of listed companies to manipulate narrative disclosure, enriching the research of the relationship between financial data and text information, and also provides a new empirical basis for the supervision department to strengthen the supervision of information disclosure and external information users to judge the quality of information.

Key words: textual description, subjectivity, earnings manipulation