Applying Transformers for Summarizing Financial Statement Data

Automatic financial statement summarization using machine learning is an important task for forecasting company performance. This paper tested 4 transformer architectures to determine their effectiveness for summarizing financial statements. Further applications of the methods presented here could include summarization of a wider variety of corporate documents, particularly for the private equity and venture capital markets.

Based on the results obtained for financial statement summarization, the PEGASUS- Legal v2 model is the optimal model with the highest ROUGE scores, with ROUGE-1 precision at 0.37 precision and an F-measure of 0.33. This model outperformed the T5, BERT2BERT and PEGASUS-Finance models for the summarization task.

Example SHAP Value for explaining how a summary was generated: