Abordagem híbrida e independente de domínio para extração de aspectos na análise de sentimentos

Abstract

Opinions are central in most of the human activities and are keys of influence to our behaviors. Our beliefs, perception of reality and our choices are in a considerable degree, influenced by how people see and evaluate the world. In view of this statement theSentimentAnalysis(SA)hasbeengrowingconstantly,thepossibilityofunderstand people’sfeelingsandopinionsaboutcertainsubjectsgetseveryoneexcited.Sentiment Analysisisthecomputationalstudyofpeople’sopinions,attitudesandemotionsabout some entity. The literature about Sentiment Analysis is pretty wide, having too many ways of execute such tasks. A variation of SA called Aspect based Sentiment Analysis (ABSA) has been receiving researchers attention. In this approach feelings are identified in relation to sentence aspects, in order to discern those that are treated in eachsentenceordocument.ABSAisdividedinthreemajortaskswhicharetheextraction,classificationandaggregationoftheaspect,havingaspectextractionasthemost complextask.There’sseveralapproachestosolvetheaspectextractiontask,although manyoftheseapproachesaredomaindependent,makingdifficulttoreplicatetheseapproaches to domains that does not have the same features. Therefore, this work aims topurposeadomainindependenthybridmethodtoaspectsextraction,thatconsistsin fourmajorsteps.Thefirstoneidentifyallthepossibleaspectsoutofsemanticrulesfor eachsentence.Afterthisstep,willbegeneratedalexicalofallthesentenceshavingthe aspectsandmostrelevantfeelings.Inthefollowstepismadethepruningofpossibleaspectsusingsemanticrulesthroughthelexicalofaspectsandfeelingsmadepreviously. Lastly,ismadeaselectionamongtheremainingaspectsbyadynamicthreshold.This purpose was evaluated in the Semeval’s dataset, containing opinions about several aspects related to restaurants and laptops, using the most adopted evaluation metrics in literature. The experimental results imply that the proposed method is competitive when it’s compared to many other methods dependents and independents of state of art domain.