Evaluation of deep learning algorithms in sentiment analysis

Deep Learning (DL) techniques have played an important role in the solution of a wide range of problems. Convolutional Neural Networks (CNN) are especially good at image processing tasks. Recurrent Neural Networks (RNN) are usually applied in Natural Language Processing (NLP) tasks. Sentiment analysis has been a popular area of research because people's opinions are important for each other and people have been sharing their opinions in social media freely. In the literature, there are studies that use simple deep learning techniques or their combinations in sentiment analysis. In this thesis, we evaluate different deep learning techniques. The learning models we compare are CNN, Long Short-Term Memory Networks (LSTM), and their ensembles and combinations. Moreover, we use Support Vector Machines (SVM) in a combination. In addition to these models, we compare different word embedding techniques such as Word2Vec and Global Vectors for Word Representation (GloVe) models. We focus on sentiment analysis from Twitter Data provided in Semantic Evaluation (SemEval), which is one of the most popular international workshops on semantic evaluation. Sentiment analysis work consists of two main steps. First phase is the creation of word embeddings, Word2Vec and GloVe models are compared in this step. Second phase is supervised training. Here, we apply CNN, LSTM, SVM, and their various combinations. Additionally, we have compared voting individual results of different learning techniques and their more organic combinations. All these combinations are tried with different weights and parameters, and best scoring values of each model is compared in terms of accuracy, precision, recall, and F-score.