Darkhan Rysbek, M.Sc.
Department of Scientific Computing
May 2019
Supervisor: Ömür Uğur (Institute of Applied Mathematics, Middle East Technical University, Ankara)
Abstract
Easier access to computers, mobile devices, and availability of the Internet have given people the opportunity to use social media more frequently and with more convenience. Social media comes in many forms, including blogs, forums, business networks, review sites, and social networks. Therefore, social media generates massive sources of information in the shape of users` views, opinions, and arguments about various products, services, social events, and politics. By well-structuring and analysing this kind of data we can obtain significant feedbacks about products and services. This area of research is typically called sentiment analysis or opinion mining. In the last decade, this field of Natural Language Processing (NLP) has witnessed a fascinating progress due to Deep Neural Network (DNNs).
Recurrent Neural Network (RNNs) are one of the main types of DNN architectures which are used at modelling units in sequence. They have been successfully used for sequence labelling and sequence prediction tasks, such as handwriting recognition, language modelling, machine translation, and sentiment analysis.
Most of the studies carried on sentiment analysis using RNNs have been focused on English texts and some researches have studied on different languages. In this thesis, sentiment classification using RNNs is applied on Turkish reviews domain. Additionally, different types of word representations are used to achieve acceptable results. This dissertation presents a description of the considered model architectures and comparison of them with various word representations on two Turkish movie reviews datasets. Generally, our experimental results show that RNN models achieve reasonably good results on Turkish texts as on English texts and choice of different word representations can improve the performance of the approaches.
Keywords: Sentiment Analysis, Natural Language Processing, Deep Neural Networks, Recurrent Neural Networks, Machine Learning, Turkish