StylePTB

我们引入了一个大规模的基准 StylePTB，其中 (1) 成对的句子经历了 21 次细粒度的风格变化，跨越了文本的原子词汇、句法、语义和主题转移，以及 (2) 允许建模的多次转移的组合细粒度的风格变化作为更复杂、更高层次的转移的基石。通过在 StylePTB 上对现有方法进行基准测试，我们发现它们难以对细粒度的更改进行建模，并且在组合多种样式时更加困难。We introduce a large-scale benchmark dataset named StylePTB. It has two core characteristics: (1) Paired sentence pairs undergo 21 fine-grained style changes, covering atomic lexical, syntactic, semantic, and thematic shifts of text; (2) It supports modeling compositional fine-grained style changes through combining multiple transfers, which serve as building blocks for more complex and high-level style shifts. By benchmarking existing state-of-the-art methods on StylePTB, we find that these methods struggle to model fine-grained modifications, and face even greater challenges when combining multiple styles.

搜集汇总

背景与挑战

背景概述

StylePTB是一个大规模文本风格转换基准数据集，包含成对句子，覆盖21种细粒度的风格变化，涉及词汇、句法、语义和主题层面，旨在评估模型处理复杂风格组合的能力。该数据集由卡内基梅隆大学于2021年发布，用于推动文本风格转换研究。