TEXREL

TEXREL数据集由ASAPP的Hugh Perkins提出，专注于新兴通信领域，特别是关系研究。该数据集规模庞大，包含100,000个训练示例，每个示例包含256张图像，旨在支持快速训练并避免过拟合。TEXREL通过使用非符号输入（如图像）来表示意义，每个图像包含一个或多个具有形状和颜色的对象，以及对象间的相对位置来表达关系，增加了额外的意义维度。此数据集不仅提供了高维度的意义空间，还通过使用低分辨率图像支持卷积网络，使得在相对低计算资源环境中易于使用，适用于探索意义空间维度对组合性的影响。

The TEXREL dataset, proposed by Hugh Perkins of ASAPP, focuses on the emerging field of communication, with a particular emphasis on relational research. This large-scale dataset comprises 100,000 training examples, each consisting of 256 images, and is developed to support rapid training and mitigate overfitting. TEXREL conveys meaning via non-symbolic inputs such as images: each image includes one or more objects with specific shapes and colors, and the relative positional relationships between these objects serve to transmit relational semantics, thereby adding an extra dimension of meaning. This dataset not only provides a high-dimensional meaning space, but also supports convolutional networks through low-resolution images, making it easily applicable in environments with relatively limited computational resources, and is well-suited for investigating the impact of meaning space dimensions on compositionality.