Sentiment Analysis

Analiza sentimenta zadatak je obrade prirodnog jezika čiji je cilj stvoriti automatizirani sustav koji je stanju odrediti sentiment tekstova pisanih prirodnim jezikom. U ovom radu provedena je analiza sentimenta na skupu tweetova pisanih hrvatskim jezikom. Za svaki je tweet u podatkovnom skupu ručno označen njegov sentiment. Tweetovi su najprije pretprocesirani. Pretprocesiranjem je iz tweetova uklonjen šum te su nakon toga oni tokenizirani i lematizirani. Zatim je provedena eksploratorna analiza korištenog skupa tweetova, a nakon toga klasifikacija sentimenta primjenom metoda strojnog učenja. Za potrebe klasifikacije sentimenta istrenirani su naivni Bayesov klasifikator, logistička regresija i stroj potpornih vektora te njihov glasački ansambl i stacking ansambl s logističkom regresijom, odnosno strojem potpornih vektora kao završnim estimatorom. Klasifikatori su trenirani na podatkovnim skupovima tweetova u kojima su emotikoni zamijenjeni odgovarajućim tekstom ili s uklonjenim emotikonima čije su značajke izdvajane metodom vreće riječi ili TF-IDF metodom. Dobiveni su zadovoljavajući rezultati. Točnosti klasifikatora kreću se od 62% do 65%, a njihove F1 mjere od 67% do 71%. Klasifikatori su postigli slične rezultate na skupovima tweetova s emotikonima i bez njih, ali nešto bolje rezultate uz TF-IDF metodu izdvajanja značajki.Sentiment analysis is a task of natural language processing whose goal is to create an automated system that is able to determine the sentiment of texts written in natural language. In this paper, the sentiment analysis was performed on a data set of tweets written in the Croatian language. For each tweet in the data set, its sentiment is manually annotated. Tweets are firstly pre-processed. As a part of pre-processing noise was removed from the tweets, after which they were tokenized and lemmatized. An exploratory analysis of the set of tweets used was then performed, followed by the sentiment classification using machine learning methods. For sentiment classification, the naive Bayes classifier, logistic regression and support vector machine, their voting ensemble and stacking ensemble with logistic regression or support vector machine as the final estimator were trained. The classifiers were trained on tweet data sets in which the emoticons were replaced with appropriate text or were removed. The features from the tweets datasets were extracted using TF-IDF and bag-of-words methods. Obtained results are satisfactory. The accuracy of the classifiers ranges from 62% to 65%, and their F1 measures from 67% to 71%. Classifiers achieved similar results on tweet data sets with and without emoticons, but slightly better results with the TF-IDF method of feature extraction