A novel hybrid and ensemble deep learning model for enhancing sentiment analysis in Persian context

Abstract

Advances in mobile technologies have established social media as a major platform for expressing emotions and opinions. Analyzing such publicly shared sentiments enables companies and political organizations to make informed and data-driven decisions. As a result, sentiment analysis has become an essential tool for researchers and policymakers seeking to accurately interpret public opinion. However, traditional text processing approaches often struggle to achieve high accuracy when dealing with complex and informal textual data, highlighting the need for more robust analytical frameworks. Although deep learning techniques have recently achieved notable success in Natural Language Processing (NLP), research on Persian sentiment analysis remains relatively limited. The primary objective of this study is to propose a novel hybrid and ensemble deep learning framework specifically designed for sentiment analysis in the Persian language. To assess the effectiveness of the proposed approach, multiple architectures were examined, including standalone models (BERT, RoBERTa, RNN, and BiLSTM), ensemble models based on averaging and majority voting strategies, and two hybrid configurations that integrate sequential and transformer-based models. Experimental results indicate that BiLSTM outperforms other standalone approaches, while the ensemble model using average voting and the hybrid configuration Hybrid-1 (BERT-RNN-BiLSTM) achieve the highest classification performance across accuracy, precision, recall, and F1-score. These findings demonstrate the effectiveness of integrating diverse deep learning strategies and emphasize the practical applicability of hybrid and ensemble architectures for Persian sentiment classification. This study presents a comprehensive and systematic investigation of hybrid transformer–recurrent architectures and ensemble learning strategies for Persian sentiment analysis, supported by extensive experiments on benchmark datasets. The proposed models were evaluated on four benchmark Persian sentiment analysis datasets, namely the Persian Electronic Dataset, MirasOpinion Dataset, Instagram Persian Comment Dataset and SentiDariPers Dataset. Experimental results, supported by comparative analysis with existing studies, demonstrate the consistent superiority of the proposed framework across all datasets. In particular, the Hybrid-1 (BERT-RNN-BiLSTM) model achieved the highest accuracy of 97.00% on the Persian Electronic Dataset, while accuracies of 84.00%, 80.75% and 95.19% were obtained on the MirasOpinion, Instagram Persian Comment and SentiDariPers datasets, respectively. Furthermore, the Hybrid-2 (RoBERTa-BiLSTM-RNN) and the proposed ensemble models also demonstrated strong and consistent performance across all four datasets, confirming the robustness and effectiveness of the proposed architectures for Persian sentiment classification.