原创力文档- 1、本文档共27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)。
ReframingHuman-AICollaborationforGeneratingFree-TextExplanations
SarahWiegreffe? JackHessel? SwabhaSwayamdipta?MarkRiedl? YejinChoi令?
?SchoolofInteractiveComputing,GeorgiaInstituteofTechnology
?AllenInstituteforArtificialIntelligence
令PaulG.AllenSchoolofComputerScienceandEngineering,UniversityofWashington
saw@,riedl@,{jackh,swabhas,yejinc}@
Abstract
Largelanguagemodelsareincreasinglycapa-bleofgeneratingfluent-appearingtextwithrel-ativelylittletask-specificsupervision.Butcanthesemodelsaccuratelyexplainclassificationdecisions?Weconsiderthetaskofgeneratingfree-textexplanationsusinghuman-writtenex-amplesinafew-shotmanner.Wefindthat
(1)authoringhigherqualitypromptsresultsinhigherqualitygenerations;and(2)surpris-ingly,inahead-to-headcomparison,crowd-workersoftenpreferexplanationsgeneratedbyGPT-3tocrowdsourcedexplanationsinexist-
“Usuallyahamburgerwithfriendsindicatesagoodtime.”
EvaluationSelect-1
Evaluation
Explanation-Level
topscoreOvergeneratingExplanationFilteringAcceptableExplanationsAcceptabilityPredictions
topscore
OvergeneratingExplanationFilteringAcceptable
Explanations
T5
T5/T0
FilterModel
CrowdworkerAcceptability
ingdatasets.Ourhumanstudiesalsoshow,however,thatwhilemodelsoftenproducefac-tual,grammatical,andsufficientexplanations,theyhaveroomtoimprovealongaxessuch
greedy
GPT-3
GPT-3DaVinci
Labels
asprovidingnovelinformationandsupportingthelabel.WecreateapipelinethatcombinesGPT-3withasupervisedfilterthatincorpo-ratesbinaryacceptabilityjudgmentsfromhu-mansintheloop.Despitetheintrinsicsubjec-tivityofacceptabilityjudgments,wedemon-stratethatacceptabilityispartiallycorrelatedwithvariousfine-grainedattributesofexplana-tions.Ourapproachisabletoconsistentlyfil-terGPT-3-generatedexplanationsdeemedac-ceptablebyhumans.
1 Introduction
Asna
文档评论（0）