Attention Alignment under Logical Constraints for Reliable Financial Statement Reasoning

Abstract

This study addresses the lack of verifiability in attention mechanisms, cross-statement inconsistencies, and cross-period attention drift during financial statement reasoning, and proposes an attention consistency verification framework grounded in financial logic. The method first encodes structured financial indicators and note information within a unified representation space, enabling the integration of multiple financial elements. It then constructs a logical constraint matrix based on accounting identities, cross-statement dependencies, and hierarchical account structures, and maps the model's raw attention distribution into a logic-compliant attention space to strengthen associations among key financial items. The framework further introduces temporal consistency measures and cross-statement consistency measures to evaluate the stability of attention behavior across reporting periods and statement types, thereby producing quantifiable indicators of reasoning consistency. An attention alignment loss is also designed to improve the semantic validity of attention distributions while preserving the model's expressive capacity under financial structural constraints. Experimental results show that the framework achieves notable improvements over existing methods across multiple performance metrics and demonstrates strong robustness through hyperparameter, environmental, and data sensitivity analyses. The findings provide methodological support for building intelligent financial analysis systems with structural transparency, logical consistency, and reliable reasoning, and they can be applied across a wide range of financial data processing and credibility assessment tasks.