Performance Analysis of Different Sentiment Polarity Dictionaries on Turkish Sentiment Detection

Abstract

In places such as social media and news websites, the number of web-based textual materials those deal with today’s events and contain people’s emotions continues to increase inevitably day by day. All of these texts, with their own importance, affect our society in some way. For this reason, it is very important to automatically detect the sentiment polarities of these texts. In this study we developed several statistical-based semantic algorithms for Turkish sentiment analysis task. Furthermore, we conducted experiments on sentiment analysis in Turkish texts using different sentiment polarity dictionaries. We perform a number of experiments on some datasets which we collected from Twitter platform. In our experimental environment we also use two dictionaries to get the sentiment polarity score of Turkish terms and phrases: We built the first sentiment polarity dictionary by using a translator and this dictionary includes about 159,876 Turkish words. We built the second sentiment polarity dictionary by using GDELT (Global Data on Events, Languages and Tone) and this dictionary includes about 84,744 Turkish words. We also implement the state of the art baseline algorithms in order to compare the performance results. There are three important outcomes of this study: 1.) We built two publicly available sentiment polarity dictionaries for Turkish, 2.) We developed statistical-based novel semantic algorithms for Turkish sentiment analysis task, 3.) We report the observations of the effect of using different semantic-polarity dictionaries on Turkish sentiment polarity detection problem. Experiment results show that the algorithms we have developed are valuable because they give higher classification performance than the baseline algorithms on Turkish sentiment polarity detection task.