Deep Learning for Sentiment and Emotion Detection in Multilingual Contexts
Analytics
668 views ◎609 downloads ⇓
Abstract
Social media is growing as a communication medium where people can express online their feelings and opinions on a variety of topics in ways they rarely do in person. Detecting sentiments and emotions in texts have gained a considerable amount of attention in the last few years. Thus, the terms sentiment analysis and emotion detection have taken their own path to become essential elements of computational linguistics and text analytics. These terms are designed to detect peoples' opinions and emotions that consist of subjective expressions across a variety of products or political decisions. Recently, the Arab region has played a significant role in international politics and in the global economy, which has grasped the attention of political and social scientists. Yet, the Arabic language has not received proper attention from modern computational linguists. This dissertation provides a comprehensive study of sentiment analysis and emotion detection on Twitter data and analyzes the existing work that has been accomplished to detect and analyze English and Arabic tweets. It also examines a case study where a random sample of tweets has been extracted that reflect people's sentiments regarding a political event. In this case study, an R package "Sentiment" has been applied to detect sentiments and emotions in the extracted tweets. The results demonstrate a need for more investigation towards improving the effectiveness and efficiency of sentiment and emotion detection systems. Therefore, the main contribution of this dissertation is to propose a system that automatically determines the intensity of sentiments and emotions in both languages. Emotion detection for Arabic text is relatively new; to the best of our knowledge, the proposed system is the first system to detect the intensity of emotions for Arabic text using deep learning approaches. The main data inputs to the system are a combination of word and document embeddings and a set of psycholinguistic features (e.g., AffectiveTweets Weka-package, Deepmoji, Unsupervised Sentiment Neurons). Our approach is novel in using and applying CNN-LSTM with fully connected neural network architecture to obtain performance results that show substantial improvements in Spearman correlation scores over the baseline models. In addition to the aforementioned contributions, this dissertation aims to optimize the model performance for both languages by constructing and selecting informative features. It illustrates the contribution of deep learning in sentiment and emotion detection and highlights the role of using the extracted features from raw Arabic tweets and Arabic tweets translated into English during this process.