Sentiment Analysis in Turkish Using Transformer-Based Deep Learning Models

Social media, e-commerce, review, and blogging websites have become important sources of knowledge as information and communication technology has advanced. Individuals can share their thoughts, complaints, feelings, and views on a wide range of topics. Because it seeks to identify the orientation of the sentiment present in source materials, sentiment analysis is a key field of research in natural language processing. Sentiment analysis is a natural language processing (NLP) task that received the attention of many researchers and practitioners. The majority of earlier studies in sentiment analysis mainly focused on traditional machine learning (i.e., shallow learning) and, to some extent, deep learning algorithms. Recently, transformer-based models have been developed and applied in different application domains. These models have been shown to have a huge potential to advance text classification and, particularly, sentiment analysis research fields. In this paper, we investigate the performance of transformer-based sentiment analysis models. The case study has been performed on four datasets that are in Turkish. First, preprocessing methods were used to remove links, numerals, unmeaningful, and punctuation characters from the data. Unsuitable data was eliminated after the preprocessing phase. Second, each data set splitted into two parts; 80% for training, 20% for testing. Finally, transformer-based BERT, ConvBERT, ELECTRA, traditional deep learning, and machine learning algorithms have been applied to classify sentences into two or three classes, which are either positive, neutral, or negative. Experimental results demonstrated that transformer-based models could provide superior performance in terms of F-score compared to the traditional machine learning-based and deep learning models.