Extraction of characteristic information from financial super-long texts and prediction of corporate violations

Annual report texts contain clues about corporate misconduct. Predicting misconduct through AI-based analysis of these texts can help investors better avoid risks. However, due to the current limitations of AI language models, embedding the semantic vectors of long text paragraphs from annual reports faces a trade-off between "globality" and "accuracy." By using machine learning models (DecisionTree, RandomForest, LightGBM), our study compares the effectiveness of annual report text information at four segmentation granularities in predicting corporate misconduct. We find that, with single-granularity encoding, the Bert-Sentence-Stack semantic extraction method provides more effective annual report text encodings for predicting misconduct, achieving a best AUC of 0.7250. Furthermore, by implementing multi-granularity feature fusion, we achieve a winning combination of "globality" and "accuracy" with a maximum AUC of 0.7701. Compared to using financial features alone, multi-granularity text feature fusion increases the prediction AUC for corporate misconduct by about 12 %, indicating that multi-granularity text semantic features provide valuable incremental information. This study offers new insights and solutions for the integration and utilization of long financial texts and information mining.