Dataset Card for StructBill-CN

StructBill-CN is a comprehensive benchmark dataset tailored for Schema-based Unified Extraction in Visual Document Understanding (VDU). It specifically targets the direct-ingestion extraction of complex, hierarchical information (both global Key-Value pairs and nested wireless tables) from high-resolution Chinese medical statement images, with a strong emphasis on evaluating structural accuracy and arithmetic logical consistency.

Dataset Details

Dataset Description

Automated transformation of complex statement images into queryable databases is a critical yet unresolved challenge. While Multimodal Large Language Models (MLLMs) excel in general perception, they struggle with precise direct-ingestion tasks, particularly when processing wireless tables where the absence of visual grid lines renders traditional Table Structure Recognition (TSR) ineffective.

StructBill-CN bridges the gap between visual cues and semantic structure. Unlike traditional datasets that focus heavily on physical bounding boxes, StructBill-CN features logical structure annotations for both global Key-Value pairs and complex line-item tables. It compels models to comprehend semantic layouts and business logic (such as deterministic arithmetic rules like Price * Quantity = Amount) rather than merely performing physical visual detection.

• Curated by: The authors of the paper "StructBill-CN: Benchmarking and Improving Logical Consistency in Visual Document Understanding with Schema-Reinforced Policy Optimization"
• Language(s) (NLP): Chinese (zh-CN)
• License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)

Dataset Sources

To strictly comply with original data distribution agreements, our repository explicitly separates our novel annotations from third-party raw images.

• Original Images - CHIP-2022: Aliyun Tianchi Platform - CHIP 2022 Shared Task (Requires Tianchi account)
• Original Images - SIBR-med: Official SIBR-med GitHub Repository
• Paper: Under Review at IJCAI 2026

Uses

Direct Use

This dataset is designed for academic research in the field of Document AI, specifically for:

• Evaluating and training Multimodal Large Language Models (MLLMs) on Visual Information Extraction (VIE).
• Benchmarking models on parsing complex, wireless, and borderless tables.
• Assessing the logical reasoning and arithmetic consistency capabilities of VDU systems.
• Developing reinforcement learning algorithms (like SRPO) for document alignment and schema-following tasks.

Out-of-Scope Use

• Commercial Use: Prohibited under the CC BY-NC-SA 4.0 license.
• High-Risk Decision Making: Models trained on this dataset should not be deployed in real-world healthcare or financial automated auditing without a robust Human-in-the-loop (HITL) review system.

Dataset Structure

StructBill-CN comprises 2,300 high-resolution images covering 6 distinct business schemas.

Strict Compliance Distribution Policy (Decoupled Release):

To comply with data privacy policies and third-party distribution agreements, the dataset is structured as follows:

• 1. What We Provide Here (Available Now):

  • Unified Annotations: Our curated, hierarchical JSON annotations for all 3,596 instances (CHIP-2022, SIBR-med, and Internal-Wild).

  • Internal-Wild Test Images: The original image files for the Out-of-Distribution (OOD) test set of our proprietary Internal-Wild data. These have been fully de-identified.

• 2. Third-Party Source Images (Download Required):

  • For the CHIP-2022 and SIBR-med subsets, we only provide the annotations. Researchers must download the original raw images directly from their respective official platforms (linked in the Dataset Sources section above) to pair with our JSON files.

Dataset Creation

Curation Rationale

Existing benchmarks predominantly focus on simple KV extraction or ruled tables with relatively static layouts. They fail to expose model deficiencies in semantic alignment when dealing with borderless tables, structural ambiguity, and extreme visual density. StructBill-CN was created to establish an "Ingestion-Ready" benchmark that mimics real-world database schemas, forcing models to infer structure from content logic.

Source Data

Data Collection and Processing

The dataset aggregates data from three main sources:

1. CHIP-2022 (1,700 items): Inpatient/Outpatient/Pharmacy invoices and Discharge records.
2. SIBR-Med (600 items): Fee lists and Notification notes.

Annotations

Annotation process

The annotation protocol strictly prioritizes semantic attribution over physical location. Instead of traditional bounding box coordinates, annotations are formatted as a Hierarchical JSON standard. In the presence of printing offsets or wireless table layouts, labels are assigned based on the logical business context. Furthermore, all numerical fields (Price, Quantity, Amount) were cross-validated to ensure arithmetic consistency in the Ground Truth.

Recommendations

Users should be aware that while the dataset aims to benchmark arithmetic consistency, current MLLMs may still hallucinate numbers. It is highly recommended to implement deterministic rule-checkers on top of model outputs when using these systems in practical scenarios.

Citation

BibTeX:

@article{structbillcn2026,
  title={StructBill-CN: Benchmarking and Improving Logical Consistency in Visual Document Understanding with Schema-Reinforced Policy Optimization},
  author={Anonymous Authors},
  journal={Under Review at IJCAI},
  year={2026}
}

(Note: The citation will be updated with author names and official publication details upon acceptance.)

Dataset Card Contact

For questions regarding the dataset, data licensing, or to request the removal of source images based on copyright claims, please contact: vanvan6992@gamil.com