Unified Multimodal Relation Extraction (UMRE)
收藏资源简介：
UMRE数据集是一个统一的跨模态关系抽取数据集，旨在同时提取文本实体和视觉对象之间的模态内和模态间关系。该数据集包含28种关系类型，拥有55021个标注的跨模态关系三元组，来源于12737个文本-图像对。UMRE数据集的构建过程包括两个阶段：首先，使用多模态大型语言模型识别和提取文本和图像数据中的潜在候选实体和对象；其次，招募教育背景良好的标注员分析图像和文本信息，识别文本实体和视觉对象之间的模态内和模态间关系。UMRE数据集旨在解决跨模态检索和多模态知识图谱构建等领域的相关问题。
The UMRE dataset is a unified cross-modal relation extraction dataset, which aims to simultaneously extract intra-modal and inter-modal relationships between textual entities and visual objects. It contains 28 types of relations, with 55,021 annotated cross-modal relation triples sourced from 12,737 text-image pairs. The construction of the UMRE dataset consists of two stages: first, large multimodal language models are employed to identify and extract potential candidate entities and objects from textual and image data; second, well-educated annotators are recruited to analyze the information from both texts and images, and recognize intra-modal and inter-modal relationships between textual entities and visual objects. The UMRE dataset is designed to address relevant problems in fields such as cross-modal retrieval and multimodal knowledge graph construction.
