Financial Report Summarization via Structure-Aware Modeling and Information Fidelity Constraints

Abstract

This study addresses the challenges of complex structural hierarchies, strong cross-section dependencies, and frequent factual drift in long document summarization of financial reports. It proposes a unified framework that integrates structural constraints with information fidelity modeling. Financial reports are divided into sections or paragraph-level structural units and encoded in a structure-aware manner. Structural position representations and structure-aware attention mechanisms are introduced so that decoding performs content aggregation and organization under global ordering and boundary constraints. In addition, source consistency and structural consistency constraints are incorporated into the training objective. These constraints guide the generated content to remain traceably aligned with source evidence and reduce the risk of numerical and statement deviations. To evaluate reliability in high-risk financial text scenarios, the study constructs a comprehensive evaluation framework that includes both summary quality and fidelity metrics. Systematic comparisons and sensitivity analyses are conducted on an open financial report summarization dataset. Model stability is further examined from the perspectives of hyperparameters, environmental perturbations, and data composition. The analysis focuses on the impact of boundaries of structural position embedding dimension, long document sample proportion, numerical noise injection intensity, and learning rate on supportability and numerical consistency. Experimental results show that the proposed framework outperforms multiple baseline methods on ROUGE-L, BERTScore, NumMatch, and Supported Rate. It also maintains more robust evidence alignment under various perturbation conditions. These results verify the effectiveness and necessity of combining structural constraints with information fidelity modeling for financial report summarization.