Improving Sentiment Analysis in Persian text based on combination of Stacked Auto-Encoder and Transformer-BiLSTM-CNN
Subject Areas : AI and RoboticsSina Dami 1 * , MohammadAli Sanagoo ye Moharrer 2
1 - Department of Computer Engineering, West Tehran Branch, Islamic Azad University, Tehran, Iran
2 - Department of Computer Engineering, West Tehran Branch, Islamic Azad University, Tehran, Iran
Keywords: Sentiment Analysis, Feature Extraction, Transformer, Stacked Auto-Encoder,
Abstract :
The expansion of the internet and the increasing amount of user-generated textual opinions on various topics have made sentiment analysis a crucial tool for understanding public sentiment towards different subjects. These insights are invaluable for businesses, policymakers, and society as a whole, but manually analyzing such a volume of data is costly and impractical. This study leverages automated and deep learning approaches by combining a Stacked Autoencoder (SAE) for feature extraction and a Transformer-BiLSTM-CNN model for sentiment classification, specifically designed for the Persian language. ParsBert, the Persian version of BERT, was used for data preprocessing. This combined approach demonstrated improved performance in key evaluation metrics such as accuracy, precision, recall, and F1 score, outperforming comparative models like Transformer-BiLSTM-CNN, SAE-LSTM, and CNN. Results on datasets including user reviews from the Taghcheh and Digikala platforms and Persian tweets affirm the effectiveness of this hybrid model.