Large language model driven transferable key information extraction mechanism for nonstandardized tables.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Extracting information from unstructured tables is challenging due to layout variations and reliance on large annotated datasets.
• Current methods struggle with direct output of structured formats like JSON, limiting scalability and generalization.

Purpose of the Study:

• To develop a novel mechanism for key information extraction from unstructured tables that overcomes existing limitations.
• To achieve high accuracy and generalization without task-specific fine-tuning.

Main Methods:

• Proposed the Large Language Model Driven Transferable Key Information Extraction Mechanism (LLM-TKIE).
• Employs text detection and recognition for content extraction from document images.
• Utilizes a Large Language Model (LLM) for semantic reasoning, completeness verification, and structured data organization.

Main Results:

• Achieved an F1-score of 80.9 and 88.85 accuracy on CORD, and 83.9 F1-score with 93.3 accuracy on SROIE without fine-tuning.
• Outperformed state-of-the-art multimodal large models by 5-8% accuracy on unlabeled customs domain datasets.
• Evaluated LLM performance across various sizes and quantization strategies for practical guidance.

Conclusions:

• LLM-TKIE demonstrates robust generalization and structural precision in key information extraction from unstructured tables.
• The proposed method offers a scalable and adaptable solution, outperforming existing approaches.
• Provides valuable insights for selecting and optimizing LLMs for information extraction tasks.