Neural Mining of PersianShort Argumentative Texts

Mohammad Yeghaneh Abkenar, Manfred Stede

Abstract: Argumentationmining(AM) isconcernedwithextractingargumentsfromtextsandclassifyingtheelements(e.g., claimandpremise)andrelationsbetweenthem,aswellascreatinganargumentativestructure. Asignificant hurdletoresearchinthisareaforthePersianlanguageisthelackofannotatedcorpora.Thispaperintroducesthe firstargument-annotatedcorpusinPersianandtherebythepossibilityofexpandingargumentationminingtothis language.Thestartingpoint istheEnglishargumentativemicrotextcorpuspart1(AMT)(PeldszusandStede,2015), andwebuiltthePersianvariantbymachinetranslationandcarefulpost-editingoftheoutput.Wecall thiscorpus Persianargumentativemicrotext (PAMT).Moreover,wepresentthefirstresultsforArgumentativeDiscourseUnit (ADU)classificationforPersian,whichisconsideredtobeoneofthemainsubtasksofargumentationmining.We determinetheADUsandtheir types(claimvs.premise)bytwomethods: (i)spancategorizationusingthedeep learningmodelofspaCyVersion3.0(aCNNmodelontopofBloomembeddingwithattention),and(ii)aneural sequencetagger.Theresultsthatweobtainwiththesecondapproacharecomparabletothoseachievedonthe samesubtaskinAMTanditsothertranslations.