POLITEREWRITE

POLITEREWRITE是由微软创建的一个专注于礼貌语言重写的创新数据集。该数据集包含从网络上收集的数千条带有负面情绪的句子及其对应的礼貌重写版本。与传统纯人工标注不同，POLITEREWRITE采用GPT-3.5与人工协作的标注模式，显著减少了人工标注的工作量。数据集分为两部分：10K黄金标准数据和100K由GPT-3.5标注的银标准数据。POLITEREWRITE旨在通过提供一个更复杂的重写任务，帮助开发能够避免言语攻击和冒犯表达的高情商生成模型，从而推动自然语言生成领域的研究。POLITEREWRITE is an innovative dataset dedicated to polite language rewriting, developed by Microsoft. It comprises thousands of sentences with negative emotional tones collected from the web, alongside their corresponding politely rewritten versions. Unlike traditional purely manual annotation workflows, POLITEREWRITE utilizes a collaborative annotation paradigm combining GPT-3.5 and human annotators, which substantially reduces the manual annotation workload. The dataset is split into two subsets: a 10K gold-standard dataset and a 100K silver-standard dataset annotated by GPT-3.5. The core goal of POLITEREWRITE is to support the development of emotionally intelligent generative models that can evade verbal attacks and offensive expressions, by offering a more sophisticated rewriting task, thus advancing research in the natural language generation domain.