Leveraging feature-level fusion representations and attentional bidirectional RNN-CNN deep models for Arabic affect analysis on Twitter

Abstract

Arabic affect analysis on Twitter avidly helps to capture the emotional states of individuals being expressed regarding many targets, such as world-level events, products, and services. It is the key to monitoring and advancing human intelligence, which impacts human decision-making processes efficiently. However, state-of-the-art models have not witnessed serious developments yet since they have just achieved an accuracy of around 54%. This inaccuracy is mainly due to the agglutination, dialectal variation, and morphological richness of the Arabic language, as well as the unique features of tweets, such as shortness, noisiness, and informal language. This paper presents an approach that tackles these challenges and then improves the performance of Arabic affect analysis on Twitter. First, we propose a novel feature-based fusion representation for Arabic tweets to capture the polysemy, semantic/syntactic information, and conveyed emotional knowledge; and to deal with Out-Of-Vocabulary (OOV) words. Second, we propose an attentional deep learning model based on Bidirectional Gated Recurrent Unit (BiGRU), Bidirectional Long Short-Term Memory (BiLSTM), and Convolution Neural Network (CNN) to effectively learn local and global features and provide a multilabel emotional classification. The experimental results indicate that our proposal outperforms twelve state-of-the-art and baseline methods with a significant improvement of 6% in accuracy.