Graph clustering approach to sentiment analysis

In this thesis, we aim at automatically predicting Turkish movie review scores using adjective clustering. We also measured the reliability of the two popular sentiment lexicons. In order to measure the agreement between these sentiment lexicons and human judgments, we designed a ranking experiment using pairwise comparisons. Then, we compared these sentiment lexicons and human judgments, and we gave results that show a moderate level of agreement between lexicons and human judgments. Furthermore, we performed adjective clustering task and singleton scoring to automatically assign scores to Turkish movie reviews. Adjective clustering reached an accuracy of 76%, singleton scoring reached an accuracy of 79%.