AgenticIE: An Adaptive Agent for Information Extraction from Complex Regulatory Documents

TL;DR

AgenticIE system extracts information from complex regulatory documents, achieving a 16% EM score improvement.

cs.CL 🔴 Advanced 2025-09-15 8 views
Gaye Colakoglu Gürkan Solmaz Jonathan Fürst
Information Extraction NLP Multilingual Regulatory Documents Machine Learning

Key Findings

Methodology

AgenticIE employs a planner-executor-responder pattern, adapting to user intent and document modality. It dynamically selects tools like OCR and translators to enhance extraction accuracy.

Key Results

  • AgenticIE achieved EM scores of 0.396 and 0.342 in KIE and QA tasks, outperforming GPT-4o by 16%.
  • In multilingual settings, AgenticIE's average SA score in open schema was 0.613, significantly better than baselines.
  • Experiments show AgenticIE excels in handling complex nested data structures.

Significance

This research offers a novel solution for extracting information from complex regulatory documents, especially in multilingual and irregular layouts. It not only improves extraction accuracy but also reduces manual intervention, facilitating automated quality control in the construction industry.

Technical Contribution

AgenticIE surpasses static and multimodal LLM baselines through dynamic tool selection and state tracking, offering greater adaptability and robustness. It showcases new engineering possibilities in handling multilingual and complex document structures.

Novelty

AgenticIE is the first system to use a planner-executor-responder pattern for information extraction in complex regulatory documents, excelling in multilingual and open schema settings.

Limitations

  • The system may encounter performance bottlenecks when handling extremely complex document structures.
  • High resource consumption requires optimization of planning and tool calls.

Future Work

Future work includes optimizing AgenticIE's resource usage to enhance sustainability for larger-scale applications and expanding to more languages and document types.

AI Executive Summary

In the construction industry, Declaration of Performance (DoP) documents are crucial for ensuring product quality and compliance. However, these documents often exist in multiple languages and complex formats, posing challenges for information extraction. Existing methods often fall short when dealing with these variable and irregular documents.

The AgenticIE system, through its planner-executor-responder pattern, dynamically adapts to user intent and document modality, significantly improving extraction accuracy. It employs OCR and translation tools to successfully handle multilingual and complex nested data structures.

Experimental results show that AgenticIE outperforms existing baselines in KIE and QA tasks, particularly in multilingual settings. Although resource-intensive, its potential in regulatory document information extraction is immense, with future improvements focusing on optimizing resource usage for better sustainability.

Deep Analysis

Background

Declaration of Performance (DoP) documents are mandated by EU regulations to specify characteristics of construction products. Despite their importance for quality control, their multilingual and complex formats make information extraction challenging.

Core Problem

Existing information extraction methods perform poorly in handling multilingual and irregularly formatted regulatory documents, hindering automation and high-precision extraction.

Innovation

The AgenticIE system uses a planner-executor-responder pattern, dynamically selecting tools and adapting to different document structures, solving the extraction challenges of multilingual and complex nested data.

Methodology

  • �� Use a planner to identify user intent and document modality
  • �� Executor calls OCR and translation tools for text extraction
  • �� Responder verifies and returns extraction results

Experiments

Experiments used a multilingual dataset with over 15K annotated entities, comparing AgenticIE with GPT-4o baselines, evaluating its EM scores in KIE and QA tasks.

Results

AgenticIE achieved EM scores of 0.396 and 0.342 in KIE and QA tasks, outperforming baseline models, particularly excelling in multilingual and open schema settings.

Applications

The system can be used for automated quality control and regulatory compliance checks in the construction industry, reducing manual intervention and increasing efficiency.

Limitations & Outlook

Despite its excellent performance in information extraction, AgenticIE's high resource consumption requires further optimization for improved sustainability.

Plain Language Accessible to non-experts

Imagine you are in a library searching for a specific book. AgenticIE acts like a smart librarian who can quickly find the book you need and translate it into a language you understand. It not only identifies the book's location but also handles different languages and complex shelf layouts.

ELI14 Explained like you're 14

Imagine you're playing a game where the goal is to find hidden treasure. AgenticIE is like a super helper that can find clues on the map and translate them into a language you understand. It handles different map formats and languages, helping you find the treasure quickly!

Glossary

Information Extraction

The process of automatically extracting structured information from text.

Used to extract key data from regulatory documents.

Planner-Executor-Responder Pattern

A system architecture that dynamically adapts to user intent and document modality.

Core architecture of the AgenticIE system.

OCR (Optical Character Recognition)

Technology that converts text in images into editable text.

Used to extract text from scanned DoP documents.

Multilingual Processing

The ability to process text in multiple languages.

A key feature of the AgenticIE system.

Exact Match Score

A metric that evaluates the precise match between extracted results and ground truth.

Used to assess AgenticIE's performance in KIE and QA tasks.

Open Questions Unanswered questions from this research

  • 1 How to enhance the system's multilingual processing capabilities without increasing resource consumption?
  • 2 How to optimize AgenticIE to adapt to more types of regulatory documents?

Applications

Immediate Applications

Construction Quality Control

AgenticIE can be used for automated quality control of construction products, reducing manual intervention and increasing efficiency.

Long-term Vision

Regulatory Compliance Checks

By expanding language and document type support, AgenticIE can be used for global regulatory compliance checks.

Abstract

Declaration of Performance (DoP) documents, mandated by EU regulation, specify characteristics of construction products, such as fire resistance and insulation. While this information is essential for quality control and reducing carbon footprints, it is not easily machine readable. Despite content requirements, DoPs exhibit significant variation in layout, schema, and format, further complicated by their multilingual nature. In this work, we propose DoP Key Information Extraction (KIE) and Question Answering (QA) as new NLP challenges. To address this challenge, we design a domain-specific AgenticIE system based on a planner-executor-corresponder pattern. For evaluation, we introduce a high-density, expert-annotated dataset of complex, multi-page regulatory documents in English and German. Unlike standard IE datasets (e.g., FUNSD, CORD) with sparse annotations, our dataset contains over 15K annotated entities, averaging over 190 annotations per document. Our agentic system outperforms static and multimodal LLM baselines, achieving Exact Match (EM) scores of 0.396 vs. 0.342 (GPT-4o, +16%) and 0.314 (GPT-4o-V, +26%) across the KIE and QA tasks. Our experimental analysis validates the benefits of the agentic system, as well as the challenging nature of our new DoP dataset.

cs.CL