Development of a Hybrid Method for Sentiment Analysis in Social Networks

The development of technology has led to changes in many habits in people's social lives. As a result of the pandemic, face-to-face communication has decreased significantly in many areas, especially education. People have started to spend time with mobile devices where they can quickly access social media and websites to socialize. Even though the pandemic is over, people are still widely using websites and social media applications to socialize. People instantly share their opinions about their experiences such as the movie they watched together or the restaurant they went to. This leads to constant data sharing through social media and websites. This data sharing leads to massive data collection. This data is of great interest to many organizations from marketing to advertising. These organizations are working to use the data shared for themselves or their sectors. With the developing artificial intelligence technologies, these studies bring added value to companies. One of the areas of study that adds this value is sentiment analysis, one of the natural language processing tasks, which is a sub-branch of artificial intelligence. In this study, experimental studies were carried out for sentiment analysis task on the dataset collected from IMDB web page. A new method called MBiGRUMCONV was proposed on the IMDB dataset with six BiGRUs and two Convolutional Neural Networks after the Word2Vec word embedding method. The best accuracy performance of this proposed method is 90.59% for the 80%-20% training-testing and 10% validation set, and 94.67% for the validation set. The proposed model was also cross-validation 3, 5, and 10-fold. The 5-fold cross-validation resulted in an accuracy of 90.67%. When all the results are evaluated together, it is seen that the proposed method performs competitively compared to the literature studies. Within the scope of the study, models were created with various deep learning, machine learning and ensemble learning methods after different text representation methods on open-source datasets collected from TripAdvisor, Rotten Tomatoes and Twitter. In order to achieve high model performance, in future studies, it is recommended to use bidirectional neural networks after text representation extraction with pre-trained BERT derivatives when building hybrid models.