- 1、本文档共36页,可阅读全部内容。
- 2、原创力文档(book118)网站文档一经付费(服务费),不意味着购买了该文档的版权,仅供个人/单位学习、研究之用,不得用于商业用途,未经授权,严禁复制、发行、汇编、翻译或者网络传播等,侵权必究。
- 3、本站所有内容均由合作方或网友上传,本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺!文档内容仅供研究参考,付费前请自行鉴别。如您付费,意味着您自己接受本站规则且自行承担风险,本站不退款、不进行额外附加服务;查看《如何避免下载的几个坑》。如果您已付费下载过本站文档,您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等,请点击“版权申诉”(推荐),也可以打举报电话:400-050-0827(电话支持时间:9:00-18:30)。
1
Retrieval-AugmentedGenerationfor
AI-GeneratedContent:ASurvey
PenghaoZhao,HailinZhang,QinhanYu,ZhengrenWang,YuntengGeng,
FangchengFu,LingYang,WentaoZhang,BinCui,Fellow,IEEE
Abstract—ThedevelopmentofArtificialIntelligenceGeneratedperformanceduetotheutilizationofnovelmodelalgorithms,
Content(AIGC)hasbeenfacilitatedbyadvancementsinmodelexplosivescaleoffoundationmodels,andmassivehigh-
4algorithms,theincreasingscaleoffoundationmodels,andthequalitydatasets.Specifically,sequence-to-sequencetaskshave
2availabilityofamplehigh-qualitydatasets.WhileAIGChas
0achievedremarkableperformance,itstillfacesseveralchallenges,transitionedfromutilizingLongShort-TermMemory(LSTM)
2suchasthedifficultyofmaintainingup-to-dateandlong-tailnetworks[12]toTransformer-basedmodels[13],andimage-
rknowledge,theriskofdataleakage,andthehighcostsassociatedgenerationtaskshaveshiftedfromGenerativeAdversarialNet-
awithtrainingandinference.Retrieval-AugmentedGenerationworks(GANs)[14]toLatentDiffusionModels(LDMs)[10]
(RAG)hasrecentlyemergedasaparadigmtoaddresssuchchal-aswell.Notably,thearchitectureoffoundationmodels,ini-
Mlenges.Inparticular,RAGintroducestheinformationretrieval
7process,whichenhancesthegenerationprocessbyretrievingtiallyconstitutedbymillionsofparameters[15],[16
文档评论(0)