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

Extracting key information from unstructured tables poses significant challenges due to layout variability, dependence on large annotated datasets, and inability of existing methods to directly output structured formats like JSON. These limitations hinder scalability and generalization to unseen document formats. We propose the Large Language Model Driven Transferable Key Information Extraction Mechanism (LLM-TKIE), which employs text detection to identify relevant regions in document images, followed by text recognition to extract content. An LLM then performs semantic reasoning, including completeness verification and key information extraction, before organizing data into structured formats. Without fine-tuning, LLM-TKIE achieves an F1-score of 80.9 and tree edit distance-based accuracy of 88.85 on CORD, and an F1-score of 83.9 with 93.3 accuracy on SROIE, demonstrating robust generalization and structural precision. Notably, our method significantly outperforms state-of-the-art multimodal large models on unlabeled customs domain datasets by 5-8% in accuracy. Additionally, our evaluation of multiple large language models of various sizes across 15 quantization strategies provides valuable insights for selecting and optimizing LLMs for key information extraction tasks, offering practical guidance for system development.