MIT-10M
收藏资源简介:
MIT-10M是一个大规模的多语言图像翻译平行语料库,包含超过1000万张图像-文本对,来源于真实世界的数据,经过广泛的数据清洗和多语言翻译验证。数据集包含84万张图像,分为28个类别,涵盖14种语言的图像-文本对,任务难度分为三个级别。数据集的创建过程包括数据收集与预处理、OCR标注与清洗、多语言翻译与验证三个主要阶段。MIT-10M旨在解决现有数据集在规模、多样性和质量上的不足,特别适用于评估和训练多语言图像翻译模型,提升模型在复杂场景下的适应性和性能。
MIT-10M is a large-scale multilingual image-text parallel corpus containing over 10 million image-text pairs sourced from real-world data, which has undergone extensive data cleaning and multilingual translation validation. The dataset includes 840,000 images divided into 28 categories, covering image-text pairs in 14 languages, with tasks divided into three difficulty levels. The creation process of MIT-10M consists of three main stages: data collection and preprocessing, OCR annotation and cleaning, as well as multilingual translation and verification. MIT-10M aims to address the shortcomings of existing datasets in terms of scale, diversity and quality, and is particularly suitable for evaluating and training multilingual image-text translation models to improve their adaptability and performance in complex scenarios.

- 1MIT-10M: A Large Scale Parallel Corpus of Multilingual Image Translation清华大学软件学院, 天津大学智能与计算学院, 百度公司 · 2024年



