A Hybrid Deep Learning Network for Sentiment Analysis on SemEval-2017 Dataset

Nowadays, the vast amount of information on the internet enables people to be informed of other thoughts and ideas quickly and get an insight to use in their decisions. Knowing how people feel about a person or event can significantly impact the decisions of individuals and organizations. With the advent of social networks and their high popularity, most people tend to share their opinions on the internet. Analyzing these feelings on social networks, as a good representation of society, can help make organizational decisions and forecast important events. Therefore, processing large volumes of information is a challenge that many researchers have considered. This study aims to present a new approach to analyzing emotions and detecting the polarity of the opinions of Twitter social network users using deep learning algorithms. The use of deep learning networks on textual data requires pre-processing and text conversion into vector space, so textual data will be transformed into vector space using the word embedding structure. This study analyzes the Twitter social network due to its data availability and significant application in emotion classification and uses the SemEval-2017 Task 4 dataset. We propose a hybrid model of the LSTM, CNN, and GRU networks, using CNN to extract the text features and the LSTM and GRU networks to preserve long-term dependencies. Moreover, we handle an imbalanced dataset using data augmentation techniques. Then we evaluate the performance of the proposed model using multiple metrics. The results show that our proposed method is about 10% better than previous related works.