Deep Reinforcement Learning-Based Early Prediction for Arabic Sentiment Analysis
Mohamed Zouidine, Hicham Hammouchi, Mohammed Khalil · 2024
In this paper, we propose a new method for Arabic sentiment analysis called early prediction. This method allows a deep learning model to stop reading tokens when it has already obtained sufficient information to make a good prediction without reading the entire input sentence. Classical deep learning models have to read the whole input, token by token, which makes them slow when it comes to classifying long sequences. The proposed method is a Bi-LSTM that learns where to stop reading and move on to the prediction step, which guarantees fast sentiment classification even with long sequences. We use reinforcement learning to train our proposed model and experiment with the Large Arabic Book Reviews dataset. Results show that the proposed method achieves fast prediction up to 16 seconds faster while maintaining a comparable accuracy when compared to the classical Bi-LSTM model.