Stacked LSTM-GRU with Double Attention model for Arabic Aspect-Based Sentiment Analysis
Athere Alghamdi, Mounira Taileb, Nada Almani · 2025
The proliferation of online opinions, fueled by the rise of social media and review platforms, has underscored the significance of sentiment analysis across diverse domains, including business, marketing, and politics. Sentiment analysis, a pivotal task within natural language processing, aims to analyze and classify sentiments expressed in textual data. Recent research has increasingly concentrated on aspect-level sentiment analysis, which offers enhanced accuracy compared to document- and sentence-level approaches by effectively addressing the multi-faceted aspects and varying sentiments present in individual reviews. However, implementing aspect-based sentiment analysis in the Arabic language poses unique challenges due to its intricate morphology and dialectal variations, resulting in a relative scarcity of Arabic-specific studies. This research introduces an Aspect-Based Sentiment Analysis model specifically designed for Arabic, utilizing AraBERT for word embeddings in conjunction with stacked LSTM and GRU layers to capture sequential dependencies and aspect-specific sentiments. A dual attention mechanism is employed to highlight the most pertinent words associated with the target aspect. The model's outputs are subsequently processed through a dense layer with softmax activation for sentiment classification. Experimental findings indicate that the model outperforms the performance of the state-of-the-art models, attaining an accuracy of 94.71%