This study utilizes deep learning in natural language processing to predict financial distress by analyzing concealed qualitative information in corporate annual reports. Word2Vec and BERT are employed to convert financial reports into vectors. Weighted Word2Vec and BERT-sentence are also utilized for enhanced text processing and report quantification. Experimental results based on a dataset of 62,312 Chinese listed companies from 2000 to 2021 show that weighted Word2Vec achieves an average prediction accuracy of 85.27% in cross-validation and 84.67% in sliding-time window validation. The findings indicate that incorporating semantic information from Management Discussion and Analysis (MD&A) significantly improves the performance of distress prediction models for listed companies, regardless of the text processing technique used. Text-based features become comparable to financial indicators and even surpass them as the prediction horizon extends. Combination features offer greater enhancement than financial indicators, especially for longer prediction horizons. Our study offers valuable insights and implications for key stakeholders, including investors and creditors.