- 1、本文档共29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)。
计算机研究与发展DOI:10.7544/issn1000-1239.202330801
JournalofComputerResearchandDevelopment61(5):1128?1142,2024
GPT系列大语言模型在自然语言处理任务中的鲁棒性
陈炫婷1叶俊杰1祖璨1许诺1桂韬2张奇1
1(复旦大学计算机科学技术学院上海200433)2(复旦大学现代语言学研究院上海200433)
(xuantingchen21@)
RobustnessofGPTLargeLanguageModelsonNaturalLanguageProcessingTasks
ChenXuanting1,YeJunjie1,ZuCan1,XuNuo1,GuiTao2,andZhangQi11(SchoolofComputerScience,FudanUniversity,Shanghai200433)
2(InstituteofModernLanguagesandLinguistics,FudanUniversity,Shanghai200433)
AbstractTheGPTmodelshavedemonstratedimpressiveperformanceinvariousnaturallanguageprocessing(NLP)tasks.However,theirrobustnessandabilitiestohandlevariouscomplexitiesoftheopenworldhavenotyettobewellexplored,whichisespeciallycrucialinassessingthestabilityofmodelsandisakeyaspectoftrustworthyAI.Inthisstudy,weperformacomprehensiveexperimentalanalysisofGPT-3andGPT-3.5seriesmodels,exploringtheirperformanceandrobustnessusing15datasets(about147000originaltestsamples)with61robustprobingtransformationsfromTextFlintcovering9popularNLPtasks.Additionally,weanalyzethemodel’srobustnessacrossdifferenttransformationlevels,includingcharacter,word,andsentence.OurfindingsrevealthatwhileGPTmodelsexhibitcompetitiveperformanceincertaintaskslikesentimentanalysis,semanticmatching,andreadingcomprehension,theyexhibitsevereconfusionregardinginformationextractiontasks.Forinstance,GPTmodelsexhibitsevereconfusioninrelationextractionandevenexhibit“hallucination”phenomena.Moreover,theyexperience
文档评论(0)