Genre and emotion classification by support of N-stage latent Dirichlet allocation

Classification of news headlines and social media based emotions are of great importance in terms of their use in the media sector, along with developing technology. In addition to understanding what kind of news is, you can find the type of news that the person has visited and can show ads that can attract that person's interest. It is also important to automatically classify news for news agencies. Social media has become more effective in many areas beyond communication. Users can share their feelings, thoughts and experiences with an event in social media tools like Facebook, Twitter, Blog. These tools are also used to share news and organize events. Information about the person can also be obtained through sharing in the social media. The mood of the person is estimated by going out of the shared feelings. Thus, personalized pages can be offered. The purpose of the study was to determine the types of news and the feelings of the tweets shared on Twitter. As a method, the subject modeling algorithm has been developed and used in a N-stage structure of the Latent Dirichlet Allocation (LDA). The data set created for the news was created by using sites like Milliyet, Mynet. The dataset used in the detection of the feelings of tweets was also created from Turkish tweets. While the news dataset has a maximum of 7 classifications; the dataset consists of five classes: angry, fear, happy, sad and confused. When modeling the system, the words are rooted. In the process of removing the roots, Zemberek, Snowball and the first 5 characters of words are used. For the news and tweets data sets, the classical LDA method was used first for the news, the emotion assignment for the tweets, and then a success by comparing with the actual label values. It has been observed that success has been achieved by using the GDA method, which was gradually changed with reference to the classical GDA, by performing the subject and feeling determination process from the other. Using the N-stage GDA method, the performance of the system with classifiers was measured using machine learning methods such as Naive Bayes, Multinomial Naive Bayes, Support Vector Machines, Random Forest and Multilayer Perceptron, using features extracted for each news and tweets.