Emotion Analysis in Arabic Language Applying Transfer Learning

Abstract

Applications of deep learning depend mainly on large training sets. In some cases a large data set is hardly found due to domain data rarity, data collection constraints and cost, or complicated data preparation and preprocessing. Deep transfer learning aims at exploiting the knowledge learned by a deep learning architecture to conduct another task in another domain. The problem of insufficient data for training in a target domain is manipulated by taking advantage of similar data, data distribution, network or task from a source domain. Both domains need to be related however they are different. Arabic emotion detection (ED) is a field characterized by lack of resources and available training sets are small for a deep learning implementation. On the other hand, applied deep Convolutional Neural Networks (CNNs) for sentiment analysis (SA) have reported enhanced results over the state-of-the-art. In this paper a deep CNN model is proposed for Arabic sentiment analysis based on character level representation. In addition, the model is used for applying Transfer Learning (TL) between sentiment analysis and emotion detection domains in Arabic language. The application intended to use Arabic character level features representation learned from a large sentiment data set to apply Arabic emotion detection. The implementation showed more enhanced performance in ED in terms of the accuracy measure.