Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neurosymbolic Information Extraction from Transactional Documents

Arthur Hemmer, Mickaël Coustaty, Nicola Bartolo, Jean-Marc Ogier

TL;DR

This work tackles information extraction from transactional documents by marrying neural generation with symbolic, schema-driven validation. A neurosymbolic pipeline uses LLMs to produce candidate extractions that are then filtered through syntactic, task, and domain-level checks, enforcing domain arithmetic integrity. The authors introduce a comprehensive 53-field transactional schema and relabeled datasets (CORD_TD and SROIE_TD) to support high-quality label generation for knowledge distillation. Results show that multi-layer validation improves $F_1$-scores and accuracy, especially in zero-shot and distillation settings, and the relabeled data provide a valuable benchmark for future research in structured IE on documents.

Abstract

This paper presents a neurosymbolic framework for information extraction from documents, evaluated on transactional documents. We introduce a schema-based approach that integrates symbolic validation methods to enable more effective zero-shot output and knowledge distillation. The methodology uses language models to generate candidate extractions, which are then filtered through syntactic-, task-, and domain-level validation to ensure adherence to domain-specific arithmetic constraints. Our contributions include a comprehensive schema for transactional documents, relabeled datasets, and an approach for generating high-quality labels for knowledge distillation. Experimental results demonstrate significant improvements in $F_1$-scores and accuracy, highlighting the effectiveness of neurosymbolic validation in transactional document processing.

Neurosymbolic Information Extraction from Transactional Documents

TL;DR

This work tackles information extraction from transactional documents by marrying neural generation with symbolic, schema-driven validation. A neurosymbolic pipeline uses LLMs to produce candidate extractions that are then filtered through syntactic, task, and domain-level checks, enforcing domain arithmetic integrity. The authors introduce a comprehensive 53-field transactional schema and relabeled datasets (CORD_TD and SROIE_TD) to support high-quality label generation for knowledge distillation. Results show that multi-layer validation improves -scores and accuracy, especially in zero-shot and distillation settings, and the relabeled data provide a valuable benchmark for future research in structured IE on documents.

Abstract

This paper presents a neurosymbolic framework for information extraction from documents, evaluated on transactional documents. We introduce a schema-based approach that integrates symbolic validation methods to enable more effective zero-shot output and knowledge distillation. The methodology uses language models to generate candidate extractions, which are then filtered through syntactic-, task-, and domain-level validation to ensure adherence to domain-specific arithmetic constraints. Our contributions include a comprehensive schema for transactional documents, relabeled datasets, and an approach for generating high-quality labels for knowledge distillation. Experimental results demonstrate significant improvements in -scores and accuracy, highlighting the effectiveness of neurosymbolic validation in transactional document processing.

Paper Structure

This paper contains 26 sections, 5 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neurosymbolic Information Extraction
  4. Structured Information Extraction
  5. Generative IE
  6. Validation
  7. Syntactic
  8. Task
  9. Domain
  10. Transactional Documents
  11. Schema
  12. Inference
  13. Evaluation
  14. Annotation
  15. Schema Limitations
  16. ...and 11 more sections

Figures (2)

  • Figure 1: Overview of the information extraction pipeline, where a prompt, document, and schema are processed by an LLM. Outputs are filtered for syntax, task, and domain validity, ensuring only high-quality labels are retained
  • Figure 2: Overview of the Schema we introduce for Domain-level validation of the information extracted from transactional documents.