Department of Sociology2024-11-102022979-10-95546-72-6N/A2-s2.0-85136573341https://hdl.handle.net/20.500.14288/15905Despite the importanceofunderstandingcausality, corporaaddressingcausal relationsare limited. There isadiscrepancy betweenexistingannotationguidelinesofeventcausalityandconventionalcausalitycorporathat focusmoreonlinguistics. Manyguidelinesrestrict themselvestoincludeonlyexplicit relationsorclause-basedarguments. Therefore,weproposean annotationschemaforeventcausalitythataddressestheseconcerns.Weannotated3,559eventsentencesfromprotestevent newswithlabelsonwhether itcontainscausal relationsornot. OurcorpusisknownastheCausalNewsCorpus(CNC).A neuralnetworkbuiltuponastate-of-the-artpre-trainedlanguagemodelperformedwellwith81.20%F1scoreontest set, and83.46%in5-foldscross-validation. CNCistransferableacrosstwoexternalcorpora:CausalTimeBank(CTB)andPenn DiscourseTreebank(PDTB).Leveragingeachoftheseexternaldatasetsfortraining,weachieveduptoapproximately64%F1 ontheCNCtestsetwithoutadditionalfine-tuning. CNCalsoservedasaneffectivetrainingandpre-trainingdataset for the twoexternalcorpora. Lastly,wedemonstratethedifficultyofourtasktothelaymaninacrowd-sourcedannotationexercise. Ourannotatedcorpusispubliclyavailable,providingavaluableresourceforcausaltextminingresearchers.Computer ScienceInterdisciplinary applicationsLinguisticsConference proceeding889371702043N/A3650