Selective Reading for Arabic Sentiment Analysis
Mohamed Zouidine, Mohammed Khalil · IEEE Access · 2025
In this work, we introduce a novel deep-learning method for Arabic sentiment analysis, contending that reading the entire input sequence is not always necessary. Many reviews can often be accurately classified without the need for all of the input tokens. Our method employs a reinforcement learning agent trained to select tokens with relevant information via a selection policy network. Instead of predicting sentiment polarity using the entire input, we focus only on tokens selected by the policy network. To empirically evaluate our proposed method, we conducted experiments on three Arabic sentiment analysis datasets, Large Arabic Book Reviews (LABR), Hotels Arabic Reviews Data (HARD), and Arabic Sentiment Tweets Dataset (ASTD). The results illustrate a significant improvement in Arabic sentiment classification when using our selective reading method, reaching state-of-the-art accuracy while using only a fraction of tokens. However, the method introduces computational cost due to the reinforcement learning component, and its scalability to very large datasets might require additional optimization.