3LF
收藏资源简介:
3LF是由首尔大学研究团队构建的一个三层次正式性谱系数据集,旨在解决传统形式转换任务中的监督错位问题。该数据集包含4500条精心对齐的句子三元组,分别对应非正式、随意和正式三种语体层次,其数据源自GYAFC基准语料库,并经过基于GPT-4o的大语言模型辅助重写与人工修订流程。构建过程以'随意'语体作为风格锚点,通过分解大型风格转换任务为两个更可控的过渡阶段,确保了语义对齐与风格清晰度。该数据集主要应用于可控文本生成领域,特别针对形式转换任务,旨在通过提供理论驱动的、对齐人类感知的监督信号,提升模型生成真实正式语言的能力。
3LF is a three-level formality spectrum dataset constructed by a research team from Seoul National University, aiming to address the supervision misalignment issue in traditional formality transfer tasks. This dataset contains 4500 meticulously aligned sentence triplets corresponding to three stylistic levels: informal, casual, and formal. Its data is sourced from the GYAFC benchmark corpus, and has undergone a workflow of GPT-4o-based large language model-assisted rewriting and manual revision. The construction process takes "casual" style as the stylistic anchor, decomposes the large-scale style transfer task into two more controllable transition stages, ensuring semantic alignment and stylistic clarity. This dataset is primarily applied in the field of controllable text generation, particularly for formality transfer tasks, aiming to improve models' capability to generate authentic formal language by providing theory-driven, human perception-aligned supervision signals.
数据集名称
3LF: Three-Level Formality Transfer Dataset
数据集概述
3LF 是一个用于三级语体正式度迁移的数据集,包含了从 GYAFC 语料库示例派生出的新创建的非正式和正式改写句子。
发布信息
- 相关论文链接:https://arxiv.org/abs/2605.29365
- 为尊重 GYAFC 的原始发布条件,本发布仅包含新创建的非正式和正式改写句子,不重新分发 GYAFC 的原始休闲句子。
数据集结构
informal:改写后的非正式句子formal:改写后的正式句子
数据集规模
- 1K < n < 10K
语言
- 英语(en)
标签
- formality(语体正式度)
许可证
- 其他(license: other)
引用
若在研究中使用了 3LF 数据集,请引用以下文献: bibtex @misc{yu2026casualanchorresolvingsupervision, title={Casual as an Anchor: Resolving Supervision Misalignment in Formality Transfer Dataset}, author={Hyojeong Yu and Hyukhun Koh and Minsung Kim and Kyomin Jung}, year={2026}, eprint={2605.29365}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2605.29365} }
致谢
3LF 数据集基于 GYAFC 语料库的示例构建。用户在使用本数据集时,也应引用以下文献:
- Rao, S. and Tetreault, J. (2018). Dear Sir or Madam, May I Introduce the GYAFC Dataset: Corpus, Benchmarks and Metrics for Formality Style Transfer. NAACL-HLT 2018.




