Neurosymbolic Information Extraction from Transactional Documents
Existing methods for information extraction from financial transaction documents suffer from poor zero-shot generalization and difficulty in enforcing arithmetic constraints (e.g., balance consistency, quantity conservation). Method: This paper proposes a neuro-symbolic fusion framework that leverages large language models (LLMs) to generate candidate extractions, augmented by a three-tier symbolic verification mechanism—syntactic, task-specific, and domain-specific—to rigorously enforce arithmetic consistency. It introduces the first structured schema tailored to transaction documents and enables symbolic-guided zero-shot knowledge distillation. Domain rules are tightly integrated with neural generation to enhance interpretability and cross-document generalization. Results: Evaluated on a re-annotated transaction dataset, the framework achieves significant improvements in F1 score and accuracy. Empirical results validate the effectiveness and robustness of this neuro-symbolic verification paradigm for financial document processing.