CoEdIT：通过任务特定指令调整的文本编辑

CoEdIT: Text Editing by Task-Specific Instruction Tuning

May 17, 2023

作者: Vipul Raheja, Dhruv Kumar, Ryan Koo, Dongyeop Kang

cs.AI

摘要

文本编辑或修订是人类写作过程中的一个关键功能。了解大型语言模型（LLMs）在进行高质量修订和与人类写作者合作方面的能力是朝着构建有效写作助手的关键一步。借助LLMs和指令调整的先前成功经验，我们利用经过指令调整的LLMs进行文本修订，以提高用户生成文本的质量并提高流程的效率。我们介绍了CoEdIT，这是一种用于写作辅助的最先进文本编辑模型。CoEdIT接受用户提供的指令，指定所需文本的属性，比如“使句子更简单”或“以更中性的风格写”，然后输出编辑后的文本。我们展示了一个在各种任务特定指令的多样集合上进行微调的大型语言模型（共82K个指令）。我们的模型：（1）在各种文本编辑基准测试中实现了最先进的性能，（2）与公开可用的在指令上训练的最大尺寸LLMs相比具有竞争力，同时体积减小了60倍，（3）能够推广到未见过的编辑指令，（4）具有组合理解能力，可以推广到包含不同编辑操作组合的指令。通过广泛的定性和定量分析，我们表明写作者更喜欢CoEdIT建议的编辑，相对于其他最先进的文本编辑模型。我们的代码和数据集是公开可用的。

English

Text editing or revision is an essential function of the human writing process. Understanding the capabilities of LLMs for making high-quality revisions and collaborating with human writers is a critical step toward building effective writing assistants. With the prior success of LLMs and instruction tuning, we leverage instruction-tuned LLMs for text revision to improve the quality of user-generated text and improve the efficiency of the process. We introduce CoEdIT, a state-of-the-art text editing model for writing assistance. CoEdIT takes instructions from the user specifying the attributes of the desired text, such as "Make the sentence simpler" or "Write it in a more neutral style," and outputs the edited text. We present a large language model fine-tuned on a diverse collection of task-specific instructions for text editing (a total of 82K instructions). Our model (1) achieves state-of-the-art performance on various text editing benchmarks, (2) is competitive with publicly available largest-sized LLMs trained on instructions while being sim60x smaller, (3) is capable of generalizing to unseen edit instructions, and (4) exhibits compositional comprehension abilities to generalize to instructions containing different combinations of edit actions. Through extensive qualitative and quantitative analysis, we show that writers prefer the edits suggested by CoEdIT, relative to other state-of-the-art text editing models. Our code and dataset are publicly available.