ORCHARD

ORCHARD数据集由南洋理工大学创建，旨在评估神经序列模型在多层次结构推理中的系统性泛化能力。该数据集包含500,000条序列，涉及嵌套的数学运算和单个数字整数，使用波兰表示法便于解析成语法树。每个序列包含两个句子，通过字符'X'分隔，涉及多种数学运算符如FIRST, LAST, MIN, MAX和COPY。数据集的创建过程涉及随机分配运算符和节点类型，以模拟自然语言的多样性。ORCHARD数据集主要用于解决自然语言处理中模型对多层次结构推理能力的评估问题，特别是在没有显式解析树的情况下。

The ORCHARD dataset was created by Nanyang Technological University to evaluate the systematic generalization capability of neural sequence models in multi-level structural reasoning. This dataset contains 500,000 sequences involving nested mathematical operations and single-digit integers, and adopts Polish notation to facilitate parsing into syntax trees. Each sequence includes two sentences separated by the character "X", and covers various mathematical operators such as FIRST, LAST, MIN, MAX, and COPY. The dataset construction process involves randomly assigning operators and node types to simulate the diversity of natural language. The ORCHARD dataset is primarily designed to assess models' multi-level structural reasoning abilities in natural language processing scenarios, particularly when no explicit parse trees are available.