Contributions to Improving Emotion Recognition and Sentiment Analysis From Text Data Using Deep Learning

Abstract

With the advent of Web 2.0, Affective computing conjoins the research topics of sentiment analysis and emotion recognition and has recently attracted increasing attention. It has guided computers to recognise and convey emotions and intelligently react to human sentiments and emotions. It is essential to develop a cognitive, intelligent system capable of recognising and comprehending people's sentiments and emotions and providing prompt, empathetic responses. As a result, affective analysis can avidly capture the affective aspects of individuals towards given world-level events, products, services, etc. Arabic affective analysis on social networks becomes a major concern regarding language uses, e.g., morphological richness, ambiguity, agglutination, dialectal variation and complexity, and therefore, the accurate recognition of sentiments and emotions is a more challenging task for valid and reliable affective analysis. Proposals in this context have not witnessed serious developments yet since the current state-of-the-art models achieved insignificant accuracy, whereas these models are so far unconsidered in human decision-making processes. The goal of this thesis is to study and address challenging issues in this area, focusing on both the design of efficient deep-based methods and well-distributed textual representation techniques, as well as on studying how these methods and techniques can be applied to help advance the state-of-the-art Arabic affective analysis and, hence, facilitate and support the decision-making process. First, we focus on the sentiment analysis task from Arabic text. The key objective is to overcome the limited ability of the feed-forward model by extracting unlimited contextual information by dealing with both forward and backward dependencies from the feature sequences of the Arabic sentence. In this regard, we have proposed a bidirectional LSTM deep learning-based model with the ability to extract contextual information and then predict the sentiment polarity from the Arabic text. The empirical results demonstrate the capability of the proposed deep-based model to (1) learn the context of each word in the text, (2) access both preceding and succeeding contextual features by combining a forward hidden layer and a backward hidden layer, and (3) discover richer semantic information and make full use of contextual information than LSTM. Second, we address the affect analysis problem for Arabic tweets. The main intuition behind our proposal is to exploit context-aware embedding for Arabic emotion analysis and then forward them to a deep learning model designed especially for multilabel emotion classification. In particular, we have used the transformer-based model for the Arabic language AraBERT as the semantic contextual embeddings and attention-based LSTM-BiLSTM as a multilabel emotion classification deep model. We have also enriched the tweets by transcribing their embedded emojis to their corresponding Arabic words. Then, we performed extensive experiments using different versions of AraBERT and BERT Multilingual. Finally, we focus on advancing the state-of-the-art performance of Arabic affect analysis models on Twitter. In this regard, we have proposed a high-quality representation of Arabic tweets through a four-level fusion of different features to capture the polysemy, the semantic/syntactic information, and the conveyed emotional knowledge; and to deal with out-of-vocabulary words. Next, we propose an attentional bidirectional RNN-CNN deep model that uses two parallel sequence models, namely, LSTM and GRU networks in the bidirectional layer, in order to seize high-quality context features through processing the forward and backward directions. CNN was adopted to extract meaningful features, decrease the dimensionality of feature space, and learn the local patterns. The attention mechanism was employed to enhance the essential features further. Extensive empirical experiments were carried out on benchmark and well-known datasets tailored to the Arabic language from various domains in order to demonstrate the effectiveness of our contributions. Altogether, the findings obtained during this thesis are very promising and significant enough to help academic and industrial researchers understand the recent advances, as well as new developments in this fast-paced, high-impact domain and hence, to offer an anthropomorphic path for the decision-making process.