wikiHow-TIIR
收藏资源简介:
wikiHow-TIIR数据集是基于wikiHow教程构建的,包含15万个交错式文本-图像文档的检索语料库。该数据集通过特定的管道利用大型语言模型和文本到图像生成器自动生成交错式查询。数据集在构建过程中,通过人工标注和筛选生成了7654个高质量的查询-文档对作为测试集,其余生成的查询作为训练集。该数据集旨在解决文本-图像交错式检索任务,推动相关研究的进展。
The wikiHow-TIIR dataset is constructed based on wikiHow tutorials, which encompasses a retrieval corpus of 150,000 interleaved text-image documents. It automatically generates interleaved queries via a dedicated pipeline leveraging large language models (LLMs) and text-to-image generators. During the dataset construction process, 7,654 high-quality query-document pairs were generated as the test set through manual annotation and screening, while the remaining generated queries were used as the training set. This dataset aims to address the interleaved text-image retrieval task and promote the advancement of relevant research.

- 1Towards Text-Image Interleaved Retrieval哈尔滨工业大学, 香港理工大学 · 2025年



