Deep Convolutional Arabic Sentiment Analysis with Imbalanced Data
Eslam Omara, Mervat Mosa, Nabil A. Ismail · 2019
Deep Convolutional Neural Networks (CNNs) have shown prominent performance in different NLP tasks. A basic factor in such performance is the huge amount of data used for learning. On the other hand data sets with large size, high quality and full representation needed for text analysis tasks are rarely found. In most cases existent data sets encounter class imbalance problem where data entries belong to one category is much larger than data entries represent another category. In this paper CNN performance is investigated under imbalance condition in Arabic sentiment analysis using character level representation. A data set with highly imbalanced representation of sentiment polarity classes is utilized for the investigation. Algorithm related and data related approaches are implemented to process the imbalance problem. Multiple CNN s with various configurations are tested applying cost-sensitive learning, under-sampling and oversampling techniques.