System for sentiment analysis of comments about mobile applications

Cilj diplomske naloge je bil implementirati sistem za označevanje komentarjev, ki izražajo navdušenje v spletni trgovini Google Play. Pri tem smo najprej opravili pregled področja analize sentimenta, nato pa analizirali komentarje in se bolje spoznali s problemsko domeno. Opisali smo teoretično podlago vseh metod, ki smo jih nato uporabili pri gradnji sistema. Najprej smo vhodne komentarje pretvorili v žetone besed in jih normalizirali, negi- rali in iz njih ustvarili n-grame. Nato smo uporabili korenjenje, popravljanje črkovanja, dodajanje oblikoslovnih oznak in dodatnih zunanjih atributov in ustvarili osem različnih naborov atributov. Iz vsakega nabora smo izbrali najboljše atribute s pomočjo metode χ2 . Za klasifikacijo smo nato uporabili modele, kot so naivni Bayes, logistična regresija in metoda podpornih vektorjev. Sledilo je ovrednotenje klasifikatorjev s pomočjo notranjega prečnega preverjanja, klasifikacijske točnosti, priklica, preciznosti, mere F1 in statističnih testov. Na koncu smo označevanje komentarjev iz naše problemske domene testirali na obstoječih rešitvah za analizo sentimenta in primerjali rezultate. Ugotovili smo, da obstajajo statistično pomembne razlike med rezultati klasifikatorjev. Prav tako so obstajale statistično pomembne razlike med rezultati nekaterih naborov atributov. Ugotovili smo tudi, da obstajajo statistične pomembne razlike med rezultati obstoječih rešitev in nekaterih naših modelov.The goal of this thesis was to build a sentiment analysis system, which can tag exuberant reviews in the Google Play store. First we gave an overview of the sentiment analysis field and analysis of input comments to better understand our problem domain. We described theoretical foundations of every method used to build our system. We started by transforming input reviews into tokens which were then normalized, negated and transformed in n-grams. After that we used stemming, spell correction, part of speech tagging and adding other attributes to generate eight different collections of features. We selected best features from every collection with χ2 method. For classification we used naive Bayes, logistic regression and support vector machine to classify reviews. After that we evaluated classifiers by using internal cross-validation and computing classification accuracy, recall, precision, F1 score and statistical tests. In the end we tested tagging reviews from our problem domain with existing solutions for sentiment analysis and compared the results. Results revealed that there were statistically significant differences between classifiers. There were also statistically significant differences between some feature collections. Results also revealed that there were statistically significant differences between existing solutions and some of our models