README.md

metadata

license: other
task_categories:
  - image-text-to-text
language:
  - zh
size_categories:
  - n<1K

PharmaShip: An Entity-Centric, Reading-Order-Supervised Benchmark for Chinese Pharmaceutical Shipping Documents

🔗 Paper: https://arxiv.org/abs/2512.23714

🔗 Github: https://github.com/KevinYuLei/PharmaShip

Description

PharmaShip is a real-world Chinese dataset of scanned pharmaceutical shipping documents designed to stress-test pre-trained text-layout models under noisy OCR and heterogeneous templates.

It covers three complementary tasks:

• Sequence Entity Recognition (SER)
• Relation Extraction (RE)
• Reading Order Prediction (ROP)

PharmaShip adopts an entity-centric evaluation protocol to minimize confounds across architectures and incorporates a directed acyclic reading order graph to capture layout-induced reading strategies.

Dataset Examples

PharmaShip contains scanned documents with complex tabular layouts, stamps, and handwritten text. We provide fine-grained annotations at the token, entity, and relation levels, as well as reading order supervision.

Task / View	Visualization
Token-level Annotation (Visualization of token-level ground truth)
Entity-level Annotation (Semantic entities annotated at the segment level)
Relation Extraction (RE) (Linkage between entities, e.g., Question-Answer pairs)
Reading Order Prediction (ROP) (DAG-based reading order based on relation extraction)

Dataset Statistics

PharmaShip consists of 161 annotated scanned documents with 11,295 segments. The dataset is officially split into 128 samples for training and 33 samples for validation.

Compared to existing datasets like FUNSD, CORD, and SROIE, PharmaShip features a higher density of entities and relations per sample, making it a more challenging benchmark for layout-intensive scenarios.

Table I: Statistics of PharmaShip, ROOR, FUNSD, CORD, and SROIE, including words, segments, entities, relation pairs, and the presence/strength of reading-order supervision.

Benchmark Results

We benchmarked five representative baselines: LiLT , LayoutLMv3 , GeoLayoutLM , and their RORE (Reading-Order-Relation Enhanced) variants.

The experiments demonstrate that injecting reading-order-oriented regularization consistently improves performance on SER and Entity Linking (EL) tasks.

Table II: Performance comparison of different models on SER, EL, and ROP tasks.

Note: Improvements (↑) denote F1 gains of RORE-enhanced variants. The RORE enhancement implementation is adapted from ROOR.

Access

You can load the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("YuLeiKevin/PharmaShip")

You can also download the full PharmaShip dataset from the link below:

🔗 Download PharmaShip Dataset

Note: The PharmaShip dataset can only be used for non-commercial research purpose.

Citation

If you find this dataset helpful for your research, please cite our paper:

@misc{xie2025pharmashipentitycentricreadingordersupervisedbenchmark,
      title={PharmaShip: An Entity-Centric, Reading-Order-Supervised Benchmark for Chinese Pharmaceutical Shipping Documents}, 
      author={Tingwei Xie and Tianyi Zhou and Yonghong Song},
      year={2025},
      eprint={2512.23714},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.23714}, 
}

Contact

For any questions regarding the dataset or the paper, please contact: kevinxie@stu.xjtu.edu.cn or songyh@xjtu.edu.cn, School of Software Engineering, Xi'an Jiaotong University, Xi'an, China