Geocoding location expressions in Twitter messages: A preference learning method

摘要

Resolvinglocationexpressionsintexttothecorrectphysicallocation,alsoknownasgeocodingorgrounding,iscomplicatedbythefactthatsomanyplacesaroundtheworldsharethesamename.Correctresolutionismadeevenmoredifficultwhenthereislittlecontexttodeterminewhichplaceisintended,asina140-characterTwittermessage,orwhenlocationcuesfromdifferentsourcesconflict,asmaybethecaseamongdifferentmetadatafieldsofaTwittermessage.WeusedsupervisedmachinelearningtoweighthedifferentfieldsoftheTwittermessageandthefeaturesofaworldgazetteertocreateamodelthatwillpreferthecorrectgazetteercandidatetoresolvetheextractedexpression.WeevaluatedourmodelusingtheF1measureandcomparedittosimilaralgorithms.Ourmethodachievedresultshigherthanstate-of-the-artcompetitors.