Welcome to the Visually-Rich Document(VRD) Intelligence and Understanding Kaggle competition! This competition is the sub-activity of **The Third Competition on Visually Rich Document Intelligence and Understanding (VRDIU)** in conjunction with the **39th Annual AAAI Conference on Artificial Intelligence (AAAI 2025)** in **Philadelphia, Pennsylvania, USA**. Please visit the [VRDIU competition] (https://sites.google.com/view/docui-aaai25/cfp?authuser=0) for a detailed competition schedule and important dates. ### Task Background Document parsing aims to build a structured, hierarchical understanding of visually rich documents by identifying and recognizing logical relationships between semantic entities. This structured understanding is crucial in various domains, such as academic papers, business reports, and government documents, where semantic entities like sections, subsections, tables, and figures contribute to an organized document flow. Document parsing follows document layout analysis as a downstream task, leveraging the layout structure to infer parent-child relationships between entities. For example, in an academic paper, a “Section” entity can contain multiple “Subsections,” and captions are directly associated with their corresponding figure or table bodies. In this competition, the objective is to perform document structure parsing on mining reports provided by various companies. The complexity arises as many of these reports are unstructured or scanned, presenting a unique challenge in achieving a consistent and accurate parsing outcome. Understanding the logical structure in these documents is fundamental for downstream content understanding tasks, such as key information extraction, question answering, and hierarchical organization of document content. ![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F11409544%2F364cc2972ce465d4e42958ae6dc5c91d%2Fworkshop_01.png?generation=1729726632074846&alt=media) ### Task Definition Each report comprises a collection of detected semantic entity bounding boxes in the format (x, y, w, h), where (x, y) represents the coordinates of the top-left corner, and (w, h) denotes the width and height of the semantic entity. These entities, such as *section*, *paragraph*, *table*, and *caption*, are labelled with semantic categories in the training set. The objective is to predict the logical structure of each report by establishing hierarchical relationships between these semantic entities. This involves defining how the entities are organized and nested to accurately represent the document's logical flow and structure. 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. For example, given a document with *N* detected entities, the desired output format should be a list containing the parent entity for each of the *N* entities. ***The evaluation metric is F1.*** Before conducting this competition, please ensure you read and understand the **Rules** of this competition first. Please get in touch with us anytime if you get confused about this competition. **Note a strong baseline model is provided for starting this competition!** Therefore, gather your expertise in computer vision, natural language processing, and deep learning, and join us in this competition. Engage with a vibrant community of like-minded individuals, exchange ideas, and strive for the top spot on the leaderboard. Together, let's push the boundaries of document structure parsing and shape the future of intelligent document analysis. You can scan the following QR code for more details: ![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F11409544%2Fcd2314c436cf478561bc9a645bab0083%2Fframe.jpg?generation=1731375315239589&alt=media) COMPETITION SPONSOR: Google Research If you have more questions or inquiries about this competition, feel free to fill out this competition QA form. Our organising teams will answer your questions within three working days.