A word sense disambiguation application on Turkish language using deep learning

Early symbolic approaches to understanding natural languages attempted to capture the meaning of text using rules is written by humans. However, such rule-based systems have proved to be fragile and limited in the specific areas for which they were designed. In the last 20 years, a statistical approach to natural language processing has become widespread. This system can automatically learn rules from data using mathematical models. In this thesis, deep learning studies and algorithms that can be used in word sense disambiguation task are investigated and experiments are conducted on a labeled Turkish dataset. Data from the Turkish study in SemEval-2007 is used in order to make comparisons with several previous studies. This thesis consists of five chapters. In the first chapter, a summary of word sense disambiguation is given. Then, the problem to be solved within the scope of the thesis, the purpose of solving this problem, the assumptions made about the problem and the limitations are given. In the second chapter, literature studies on word sense disambiguation studies in Turkish and foreign languages are given. In the third section, firstly, the definitions used in this study or related to this study are given. Afterwards, data collection, analysis and interpretation are given. In the section on data collection, details of the words and features used in the SemEval-2007 Turkish study are given. In the data analysis and interpretation section, firstly, the structure of the words and features used in the study and the pre-processing stages they have undergone are explained. Then, the artificial neural network models to which these data will be given as input, the training stages with these models and the analysis of the results of these stages are explained in detail. Finally, the metrics to be used in the analysis of the results of this study and their details are explained. In the fourth section, the weighted and macro averages of the results obtained using deep learning models and the metric values obtained by previous studies with the SemEval-2007 Turkish data are given and then these findings are interpreted. In the fifth section, this study is summarized in general terms and suggestions for future studies are given.