arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2406.17253cs.CL
收藏

知识编辑方法能多好地编辑令人困惑的知识?

How Well Can Knowledge Edit Methods Edit Perplexing Knowledge?

Columbia University(哥伦比亚大学);Dalhousie University(达尔豪斯大学);Stevens Institute of Technology(斯蒂文斯理工学院)

Huaizhi Ge, Frank Rudzicz, Zining Zhu

AI总结:

本文提出“困惑度”概念以衡量新知识与LLM已学概念层级的冲突程度,并构建HierarchyData数据集,通过实验证明知识困惑度与编辑有效性呈强负相关,抽象概念更难被修改。

摘要:

大型语言模型(LLMs)已展现出卓越的能力,但在训练后更新其知识仍是一项关键挑战。尽管诸如 Rank-One Model Editing (ROME) 等近期的模型编辑技术展现出潜力,但其有效性可能因被编辑知识的性质而异。我们引入了“perplexingness(困惑度)”的概念:即新知识与 LLM 已学习的概念层级和分类关系发生冲突的程度。例如,将“英国短毛猫是一种猫”编辑为“英国短毛猫是一种狗”代表了同一分类层级内的低困惑度编辑,而将“猫是一种动物”编辑为“猫是一种植物”则代表了违反基本分类边界的高困惑度编辑。为了系统地研究这一现象,我们引入了 HierarchyData,这是一个包含跨不同类别的 99 对下义词-上义词的精心策划数据集。通过在三个模型和四种编辑方法上进行的受控实验,我们证明了新知识的困惑度与知识编辑的有效性之间存在强烈的负相关性。我们的分析表明,涉及更抽象概念(上义词)的编辑通常表现出更高的困惑度,并且比其具体对应物(下义词)更难被修改。这些发现突出了 LLM 知识编辑中的一个根本挑战:一个新事实越是与 LLM 已学习的概念层级相矛盾,就越难可靠地编码该知识。

英文摘要:

Large language models (LLMs) have demonstrated remarkable capabilities, but updating their knowledge post-training remains a critical challenge. While recent model editing techniques like Rank-One Model Editing (ROME) show promise, their effectiveness may vary based on the nature of the knowledge being edited. We introduce the concept of ``perplexingness'': the degree to which new knowledge conflicts with an LLM's learned conceptual hierarchies and categorical relationships. For instance, editing ``British Shorthair is a kind of cat'' to ``British Shorthair is a kind of dog'' represents a low-perplexingness edit within the same taxonomic level, while editing ``A cat is a kind of animal'' to ``A cat is a kind of plant'' represents a high-perplexingness edit that violates fundamental categorical boundaries. To systematically investigate this phenomenon, we introduce HierarchyData, a carefully curated dataset of 99 hyponym-hypernym pairs across diverse categories. Through controlled experiments across three models and four editing methods, we demonstrate a strong negative correlation between the perplexingness of new knowledge and the effectiveness of knowledge editing. Our analysis reveals that edits involving more abstract concepts (hypernyms) generally exhibit higher perplexingness and are more resistant to modification than their specific counterparts (hyponyms). These findings highlight a fundamental challenge in LLM knowledge editing: the more a new fact contradicts an LLM's learned conceptual hierarchies, the harder it becomes to reliably encode that knowledge.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑