Automatic Extraction of Key Information from Financial Statements Using a Multi-Task Learning Model

Abstract

Financial statement information extraction and technical document data management both face significant challenges due to complex formats, scattered layouts, and intricate proximity-based semantic relations within semi-structured documents. These challenges are particularly evident in engineering enterprises involving electromagnetic devices, antenna systems, radio-frequency equipment, and related technical service activities, where accurate extraction of financial and operational information is required for reliable reporting and performance analysis. This paper proposes an automatic extraction method based on multi-task learning. The model employs a shared encoder, in which text features and layout features are fused to process four related tasks simultaneously: financial named entity recognition, key-value extraction, entity-value relationship classification, and report item classification with DistilRoBERTa. Experiments based on 500 annual reports show that the accuracy of entity recognition reaches 96.8%, the F1 score of value extraction reaches 88.7%, and the F1 score of relationship classification reaches 91.5%. The model performs particularly well in structured sections such as the balance sheet and income statement. Although its performance declines to some extent in the most complex and unstructured sections of financial statements, such as notes, the overall results demonstrate its effectiveness in improving the accuracy and semantic consistency of financial disclosure information extraction. This study provides a technical reference for intelligent document understanding and reliable data extraction in financial reporting scenarios involving engineering-oriented enterprises and electromagnetic application industries.