A Hybrid Approach for the Sentiment Analysis of Turkish Twitter Data

Abstract

Social media is now playing an important role in influencing people’s sentiments. It also helps analyze how people, particularly consumers, feel about a particular topic, product or an idea. One of the recent social media platforms that people use to express their thoughts is Twitter. Due to the fact that Turkish is an agglutinative language, its complexity makes it difficult for people to perform sentiment analysis. In this study, a sum of 13K Turkish tweets has been collected from Twitter using the Twitter API and their sentiments are being analyzed using machine learning classifiers. Random forests and support vector machines are the two kinds of classifiers that are adopted. Preprocessing methods were applied on the obtained data to remove links, numbers, punctuations and un-meaningful characters. After the preprocessing phase, unsuitable data have been removed and 10,500 out of the 13K downloaded dataset are taken as the main dataset. The datasets are classified to be either positive, negative or neutral based on their contents. The main dataset was converted to a stemmed dataset by removing stopwords, applying tokenization and also applying stemming on the dataset, respectively. A portion of 3,000 and 10,500 of the stemmed data with equal distribution from each class has been identified as the first dataset and second dataset to be used in the testing phase. Experimental results have shown that while support vector machines perform better when it comes to classifying negative and neutral stemmed data, random forests algorithm perform better in classifying positive stemmed data and thus a hybrid approach which consists of the hierarchical combination of random forest and support vector machines has also been developed and used to find the result of the data. Finally, the applied methodologies have been tested on both the first and the second dataset. It has been observed that while both support vector machines and random forest algorithms could not achieve an accuracy of up to 77% on the first and 72% on the second dataset, the developed hybrid approach achieve an accuracy of up to 86.4% and 82.8% on the first and second dataset, respectively.