Comparison of sentiment analysis with different representation models and learning methods in Turkish texts

Sentiment analysis studies have become increasingly important with the growing influence of social media in determining the prevailing sentiment in society. Individuals and organizations often use on sentiment analysis to evaluate public opinions and assess their reputation. The scope of sentiment analysis research is broad, encompassing numerous languages and offering a wide range of research areas. The advancement of technology has led to an increase in the number of methods and techniques available for sentiment analysis, highlighting the significance of comparative studies that assess these methods and techniques. While there are existing sentiment analysis studies for the Turkish language, there is a lack of sufficient research that compares different methods and classifiers, considering multiple datasets with and without preprocessing steps. In this thesis, two datasets were utilized: the KEMİK dataset, consisting of approximately 17,000 Turkish user comments, and the HUMIR dataset, which includes sub-datasets. Both preprocessed and unprocessed versions of the datasets were evaluated in the study. Various vectorization methods were employed for text classification, including the statistical approach TF-IDF, neural language models such as Word2Vec and FastText, the probabilistic model GloVe trained with neural networks, and the transformer-based model BERT. Both classical machine learning methods (SVM, LR, NB) and deep learning-based methods (RNN, CNN, LSTM, CNN LSTM) were utilized for classification. Binary and multi-class classifications were performed based on the number of classes in the datasets, and different algorithm combinations were tested. When the results obtained are compared with the success values obtained in cases where the data preprocessing steps are applied and not applied, it is seen that the data preprocessing steps do not greatly affect the success value. It has been observed that the TF-IDF method yields better results in classical machine learning techniques, while artificial neural network-based word embeddings methods perform better in deep learning-based approaches. On the other hand, when the hybrid model combinations used and the other model combinations used are compared, it is understood that the difference between the success values obtained is not very high.