Turkish Sentiment Analysis: a Comparative Study on Different Sentiment Dictionaries with Generated Features and Presenting a New Sentiment Dictionary

Sentiment analysis is a research area that aims to find out people’s opinions by matching data to topics, notions, etc. There are several approaches for sentiment analysis (e.g., machine learning-based, dictionary-based, hybrid-based, etc.). In this study, we presented a new tripolar Turkish sentiment dictionary, SentiMenTR, which consists of bigrams and unigrams. To compare the performances of SentiMenTR and other Turkish sentiment dictionaries (SWNetTR++ and SentiTurkNet), we conducted experiments on two Turkish datasets containing documents labeled as negative or positive. For experiments, firstly, we vectorized the documents by features extracted using polarity scores belonging to dictionaries. Afterward, we fitted machine learning models with these features. According to the experiment results, SentiMenTR performed better than other dictionaries. We aim to extend our dictionary, develop a negation handler module, and conduct more comprehensive experiments with deep learning methods in the future.