# AAAI-25 Visually-Rich Document(VRD-IU) Leaderboard — Data

来源：https://www.kaggle.com/competitions/aaai-25-visually-rich-document-vrd-iu-leaderboard/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

Each report consists of a collection of detected semantic entities (such as *section*, *paragraph*, *table*, *caption*, etc.), though these entities are only labelled with semantic categories in the training set. The task is to predict the logical structure by defining hierarchical relationships between the semantic entities within each report. Specifically:

**Hierarchical Relation Prediction**: Determine the parent-child relationships between document entities. For instance, if an entity “Section” is found to be the parent of a “Paragraph” entity, it indicates that the paragraph belongs to that section.

**Directional Relation:** Relationships are unidirectional, meaning if "Section" is the parent of "Paragraph," the reverse does not hold.

**Non-Bidirectional Relationships:** For each document, the aim is to construct a hierarchical tree where each entity only points upward to its parent entity, not downward to child entities.

This task requires participants to use available training data to model these logical relationships, which will help in accurate hierarchical structure prediction in the test set, where entity categories are recognized, but relationships need to be inferred.

Note: We provide the official tutorials (Data Exploration, Baseline Model) to assist participants in understanding the data structure and starting this competition easily. You can also find the tutorial links under the Code menu.
## Files

In this competition, participants can use the training dataset to develop their model, and the original validation and testing dataset will be used for leaderboard evaluation. We offer an official tutorial to help users quickly understand the dataset format and structure, making it easier for participants to get started with the competition. Link for [**dataset exploration tutorial**](https://colab.research.google.com/drive/1L3ADDnZMpPiTa_RlHgnpb_7Hdhbn5G94?usp=sharing) can be found here: (url). This dataset consists of three main splits: training, validation, and testing. For each split, we provide two types of files:


###PKL files
We've provided detailed pkl files containing essential information about document splits. The pkl file is a valuable resource as it contains essential document information, including the detected entities within each document with their responding IDs, parent-child relationship, labels, bounding boxes, textual contents inside each bounding box and the image ID of the document. This information is crucial for understanding the hierarchical structure of the visually rich documents. 


###Image Files
We provide the document images for each data split. Participants can use the information provided by the PKL file to locate the document images of each target document. Image files can provide additional layouts and visual cues to make the feature representation more comprehensive.

We also provide a default baseline using the multimodal transform where the feature extraction is based on BERT for textual feature and RoI for image feature, separately. The link can be found here:

[https://colab.research.google.com/drive/1CiQriTC7EVsEzROm_oIPTCvq208550Dz?usp=sharing](url)
[https://colab.research.google.com/drive/17QHaJ8ZaJAANkMl2lxQQ5d9dx_2gkafQ](url)
[https://colab.research.google.com/drive/1OUoq5vdGl_uAhf4E42Gx4Blcvg7L0Q-M?usp=sharing](url)
With this dataset, you can explore the challenging task of training your model to predict the parent-child relationship for each of the detected entities.. The provided files are meant to assist you in building a robust and accurate model for this competition. We hope this dataset will inspire creative solutions and foster progress in this data, Happy modelling!