Obrada negacije u kratkim neformalnim tekstovima u cilju poboljšanja klasifikacije sentimenta

Abstract

In this dissertation, the method for classifying short informal texts by sentiment was proposed. The improvement was achieved by processing the rule of syntactic negation in the Serbian language. The complexity of the grammar of the Serbian language imposes the need to systematically approach the phenomena of negation and to use the linguistic resources involved in the creation of rules for the negation treatment in its processing. The resources used are negation signals, negative quantifiers, negation intensifiers, and negation neutralizers. In addition to language resources for the application of the rules of negation, the general sentiment lexicon of positive and negative terms was used in the classification by sentiment. The evaluation of the used method was performed over a set of tweets in Serbian. Lexicon based method, as well as the supervised method of machine learning, were used for evaluation. The method presented in both cases is compared with two baseline methods: the first one that does not process the negation and the other that processes the negation, but without the rules for processing a syntactic negation. In the case where a method based on sentiment lexicon was used, the accuracy of the classification is considerably higher in relation to the two baseline methods, and the relative improvements of this method with respect to the first baseline method are the following: for the entire dataset - up to 10.62%, for a set of tweets containing negation - up to 26.63% and for a set of tweets containing negations that were processed using the rules - up to 31.16%. When using the machine learning method, higher accuracy of the classification is obtained than in the case of the lexicon-based method: for three classes - up to 69.76% and for two classes - up to 91.15%. However, the method of machine learning produces fewer improvements: for three classes up to 2.65% and for two classes up to 1.65%. The results showed a statistically significant improvement if the detected rules of negation are included in the short informal text classification method by sentiment. The results showed a statistically significant improvement if the detected rules of negation are included in the short informal text classification method by sentiment.