SynDoc: A Hybrid Discriminative-Generative Framework for Enhancing Synthetic Domain-Adaptive Document Key Information Extraction

2025-09-27

AI Summary

Vision-rich document understanding (VRDU) in specialized domains—such as medicine and finance—suffers from severe hallucination, poor domain adaptability, and heavy reliance on scarce, costly human annotations. Method: We propose a hybrid framework synergizing discriminative and generative models, integrating three key innovations: (1) structured synthetic data generation to reduce annotation dependency; (2) domain-specific query-driven adaptive instruction tuning to enhance domain knowledge injection; and (3) a recursive reasoning mechanism with multi-step verification to improve factual consistency. Contribution/Results: Extensive experiments on multiple professional document benchmarks demonstrate significant improvements in accuracy and robustness for key information extraction. The framework achieves strong generalization across domains while maintaining high domain specificity, exhibits excellent scalability, and shows strong practical viability for real-world deployment.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Domain-specific Visually Rich Document Understanding (VRDU) presents significant challenges due to the complexity and sensitivity of documents in fields such as medicine, finance, and material science. Existing Large (Multimodal) Language Models (LLMs/MLLMs) achieve promising results but face limitations such as hallucinations, inadequate domain adaptation, and reliance on extensive fine-tuning datasets. This paper introduces SynDoc, a novel framework that combines discriminative and generative models to address these challenges. SynDoc employs a robust synthetic data generation workflow, using structural information extraction and domain-specific query generation to produce high-quality annotations. Through adaptive instruction tuning, SynDoc improves the discriminative model's ability to extract domain-specific knowledge. At the same time, a recursive inferencing mechanism iteratively refines the output of both models for stable and accurate predictions. This framework demonstrates scalable, efficient, and precise document understanding and bridges the gap between domain-specific adaptation and general world knowledge for document key information extraction tasks.

Problem

• Enhancing domain-adaptive document key information extraction from complex documents
• Addressing hallucinations and inadequate domain adaptation in multimodal language models
• Bridging domain-specific adaptation with general knowledge for document understanding

Innovation

• Hybrid discriminative-generative framework for document extraction
• Synthetic data generation with domain-specific query creation
• Recursive inferencing mechanism for iterative output refinement