UniVIEDM: A Diffusion Model to Unify Visual Information Extraction Subtasks
Keywords: Visual Information Extraction
Abstract: Visual Information Extraction (VIE) focuses on extracting named entities and their relationships from visually rich document images. Traditionally, VIE systems rely on three separate models to handle three distinct subtasks, but the emerging trend in research is to design a single model that can address all of these tasks simultaneously. However, current methods face quadratic computational complexity when extracting entity relationships, as they must iterate over all token pairs. To address this issue, this paper introduces a Unified VIE Diffusion Model (UniVIEDM) for all tasks within VIE. UniVIEDM generates entity labels and their relationships conditioned on their plane coordinates, greatly reducing the computational complexity. UniVIEDM represents the layout of each visually rich document as a plane graph and converts the three subtasks into plane graph generation problems. During the pre-training stage, UniVIEDM leverages a jump-diffusion process to learn to generate valid sets of bounding boxes for all words and line segments connecting different boxes. During the fine-tuning stage, UniVIEDM employs a continuous-time Markov chain diffusion model to learn to predict the labels of boxes and line segments based on their coordinate features.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3746
Loading