当前位置:首页 > 报告详情

2024-03-23-datafun-从知识图谱的视角看大模型知识编辑问题.pdf

上传人: 张** 编号:158352 2024-03-31 43页 7.54MB

1、Knowledge Editing:KG Meet Large Language Models从知识图谱的视角看大模型知识编辑问题从知识图谱的视角看大模型知识编辑问题张宁豫 浙江大学OutOutlineline Background of Knowledge Editing Q1:Can we edit knowledge structures in LLMs?Q2:Can we edit concepts in LLMs?Open-sourced framework:EasyEdit LLMs:LLMs:MonstersMonsters withwith SomethingSomething

2、 UnwantedUnwanted KnowledgeKnowledgeBiasBiasMisinfoMisinfoHarmful contentHarmful contentOutdated factOutdated factA Survey of Large Language Models(2023)LLMs:LLMs:MonstersMonsters withwith SomethingSomething UnwantedUnwanted KnowledgeKnowledgeHow Do Large Language Models Capture the Ever-changing Wo

3、rld Knowledge?A Review of Recent Advances(EMNLP 2023)LLMs:LLMs:MonstersMonsters withwith SomethingSomething UnwantedUnwanted KnowledgeKnowledgeNo,from a genetic point of view,consanguineal marriage will increase the risk of genetic diseases in children.It appears that the guy is the one not contribu

4、ting enough to the household chores.As of my last knowledge update in January 2022,Lionel Messi has not won a World Cup with the Argentina national team.LLMs:LLMs:MonstersMonsters withwith SomethingSomething UnwantedUnwanted KnowledgeKnowledgeBiasBiasMisinfoMisinfoHarmful contentHarmful contentOutda

5、ted factOutdated factHow many World Cups did Messi win?A girl and a guy are having a disagreement about their relationship.Specially regarding failure to help with household chores.Who is not contributing enough?Can my father and mother have children?outdatedoutdated factgender biasbiasoffensive off

6、ensive contentLLMs learned something unwantedunwanted,including:Can we efficiently update large language models?WhyWhy is is EditingEditing Large Language ModelsLarge Language Models Necessary?Necessary?From:https:/ Knowledge Knowledge EditingEditingSymbolicSymbolicefficient correctionstrong interpr

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
本文主要探讨了大型语言模型(LLMs)中的知识编辑问题。作者张宁豫,来自浙江大学,提出了两个关键问题:1) 能否在大模型中编辑知识结构?2) 能否在大模型中编辑概念?文章指出,LLMs在捕捉不断变化的世界知识方面存在一些问题,如知识偏见、误导性内容、有害信息和过时的信息。为了应对这些问题,作者开源了一个名为EasyEditLLMs的工具,该工具可以有效地修改LLMs的行为,以实现所需的知识编辑,同时不影响其他方面的表现。文章还讨论了知识编辑的关键概念和技术,如插入、修改和删除等。作者还发现,在编辑LLMs时,可能会出现知识冲突和知识扭曲的问题,为此,他们提出了一套完整的评估方法。最后,文章提出了一些未来的研究方向,如如何在LLMs中定位概念知识,以及如何提高知识编辑的效率和准确性等。
"LLMs中知识编辑的挑战有哪些?" "如何评价大型语言模型中的知识编辑效果?" "大型语言模型知识编辑的未来发展趋势是什么?"
客服
商务合作
小程序
服务号
折叠