Publication:
On measuring social biases in prompt-based multi-task learning

dc.contributor.coauthorAkyürek, Afra Feyza
dc.contributor.coauthorPaik, Sejin
dc.contributor.coauthorKoçyiğit, Muhammed Yusuf
dc.contributor.coauthorAkbıyık, Seda
dc.contributor.coauthorWijaya, Derry
dc.contributor.departmentKUTTAM (Koç University Research Center for Translational Medicine)
dc.contributor.departmentGraduate School of Social Sciences and Humanities
dc.contributor.kuauthorRunyun, Şerife Leman
dc.contributor.schoolcollegeinstituteGRADUATE SCHOOL OF SOCIAL SCIENCES AND HUMANITIES
dc.contributor.schoolcollegeinstituteResearch Center
dc.date.accessioned2024-11-09T23:36:40Z
dc.date.issued2022
dc.description.abstractLarge language models trained on a mixture of NLP tasks that are converted into a textto- text format using prompts, can generalize into novel forms of language and handle novel tasks. A large body of work within prompt engineering attempts to understand the effects of input forms and prompts in achieving superior performance. We consider an alternative measure and inquire whether the way in which an input is encoded affects social biases promoted in outputs. In this paper, we study T0, a large-scale multi-task text-to-text language model trained using prompt-based learning. We consider two different forms of semantically equivalent inputs: question-answer format and premise-hypothesis format. We use an existing bias benchmark for the former BBQ (Parrish et al., 2021) and create the first bias benchmark in natural language inference BBNLI with hand-written hypotheses while also converting each benchmark into the other form. The results on two benchmarks suggest that given two different formulations of essentially the same input, T0 conspicuously acts more biased in question answering form, which is seen during training, compared to premisehypothesis form which is unlike its training examples.
dc.description.indexedbyScopus
dc.description.openaccessYES
dc.description.publisherscopeInternational
dc.description.sponsoredbyTubitakEuN/A
dc.identifier.isbn9781-9559-1776-6
dc.identifier.linkhttps://www.scopus.com/inward/record.uri?eid=2-s2.0-85137355293&partnerID=40&md5=90a1e43c3657097010282f71bbc88c72
dc.identifier.scopus2-s2.0-85137355293
dc.identifier.urihttps://hdl.handle.net/20.500.14288/12679
dc.keywordsLearning systems
dc.keywordsNatural language processing systems
dc.keywordsLanguage inference
dc.keywordsLanguage model
dc.keywordsLarge-scales
dc.keywordsMulti tasks
dc.keywordsMultitask learning
dc.keywordsNatural languages
dc.keywordsNovel task
dc.keywordsPerformance
dc.keywordsQuestion Answering
dc.keywordsText format
dc.keywordsComputational linguistics
dc.language.isoeng
dc.publisherAssociation for Computational Linguistics (ACL)
dc.relation.ispartofFindings of the Association for Computational Linguistics: NAACL 2022 - Findings
dc.subjectComputer Science
dc.subjectArtificial intelligence
dc.titleOn measuring social biases in prompt-based multi-task learning
dc.typeConference Proceeding
dspace.entity.typePublication
local.contributor.kuauthorRunyun, Şerife Leman
local.publication.orgunit1GRADUATE SCHOOL OF SOCIAL SCIENCES AND HUMANITIES
local.publication.orgunit1Research Center
local.publication.orgunit2KUTTAM (Koç University Research Center for Translational Medicine)
local.publication.orgunit2Graduate School of Social Sciences and Humanities
relation.isOrgUnitOfPublication91bbe15d-017f-446b-b102-ce755523d939
relation.isOrgUnitOfPublicatione192fff1-4efe-45a7-ab71-30233fc185a9
relation.isOrgUnitOfPublication.latestForDiscovery91bbe15d-017f-446b-b102-ce755523d939
relation.isParentOrgUnitOfPublicationc5c9bf5f-4655-411c-a602-0d68f2e2ad88
relation.isParentOrgUnitOfPublicationd437580f-9309-4ecb-864a-4af58309d287
relation.isParentOrgUnitOfPublication.latestForDiscoveryc5c9bf5f-4655-411c-a602-0d68f2e2ad88

Files