SLOG

SLOG是一个扩展自COGS的语义解析数据集，专注于结构泛化。该数据集由巴黎城市大学的李炳志等人创建，包含17个结构泛化案例，旨在评估模型对新复杂语言表达的泛化能力。SLOG通过引入新的泛化任务，如递归和填充-间隙依赖，来测试模型的结构泛化能力。数据集的生成遵循COGS的语义解析格式，使用手动指定的规则，并提供了训练和泛化集，以评估模型在未见结构上的表现。SLOG的应用领域包括评估和改进自然语言处理模型在处理复杂语言结构时的性能，特别是在模型训练中未出现的结构上的泛化能力。

SLOG is a semantic parsing dataset extended from COGS, focusing on structural generalization. This dataset was created by Bingzhi Li et al. from Paris Cité University, contains 17 structural generalization cases, and aims to evaluate models' generalization ability to novel complex linguistic expressions. SLOG tests models' structural generalization ability by introducing novel generalization tasks such as recursion and fill-gap dependencies. The dataset is generated following the semantic parsing format of COGS, using manually specified rules, and provides training and generalization splits to evaluate models' performance on unseen structures. The application scenarios of SLOG include evaluating and improving the performance of natural language processing (NLP) models when handling complex linguistic structures, especially their generalization ability to structures not encountered during model training.