Extracting Financial Subdomain Documents via An Interpretable Writing Style Pattern

We propose a novel document-level event extraction method that leverages the Spatial Fine-grained Role Writing Style, which captures the positional relationship between fine-grained entities or arguments and financial event types. This method addresses the limitation of existing methods that ignore this spatial occurrence pattern and suffer performance degradation on financial subdomain documents (such as equity pledges and repurchases). We design a combined neural model that applies this method to extract structural event information from such documents. We assess our model’s performance on various financial event extraction tasks and demonstrate its superior performance and interpretability compared to other methods. We also use linear surrogate models to analyze our model and show that it benefits more from finance-related entities in a classification pretext task, indicating that it effectively learns the spatial occurrence pattern.