BLA

BLA数据集是由阿姆斯特丹大学的研究人员创建，旨在评估多模态模型处理基本语言结构的能力。该数据集包含三种语言结构：主动-被动语态、并列和关系从句，这些结构连学龄前儿童都能掌握。数据集通过自动构建，使用Visual Genome数据集中的图像和模板生成的句子。BLA数据集不仅用于评估模型，还用于提升模型处理基本语言结构的能力，特别是在零样本学习和特定任务学习场景中。

The BLA Dataset was created by researchers at the University of Amsterdam to evaluate the ability of multimodal models to process basic linguistic structures. The dataset includes three types of linguistic structures: active-passive voice, coordination, and relative clauses, all of which even preschool children can grasp. It was constructed automatically, utilizing images from the Visual Genome dataset and sentences generated from predefined templates. The BLA Dataset not only serves as an evaluation benchmark for models but also aids in improving models' capability to handle basic linguistic structures, particularly in zero-shot learning and task-specific learning scenarios.