Quantum-inspired interpretable deep learning architecture for text sentiment analysis

Text has become the dominant medium of communication on social media, carrying rich emotional signals. Extracting such information is thus crucial for understanding public opinion and user behavior. While existing sentiment analysis approaches have achieved notable progress, they still struggle to integrate diverse semantic cues and often lack interpretability. To overcome these limitations, we introduce a quantum-inspired deep learning architecture that integrates fundamental principles of quantum mechanics (QM) with neural models for sentiment analysis. Specifically, we exploit the structural parallels between text representation and QM to devise a quantum-inspired representation scheme, further extended with a novel embedding layer. A feature extraction module based on long short-term memory (LSTM) networks and self-attention mechanisms (SAMs) captures contextual dependencies, while a density matrix formulation grounded in complex-valued quantum states enables expressive and interpretable modeling. Finally, a 2D convolutional neural network (CNN) is employed for feature condensation and dimensionality reduction. Extensive visualization, comparative, and ablation experiments demonstrate that our model not only achieves superior accuracy and efficiency over existing approaches but also provides enhanced interpretability through the integration of QM principles. The implementation is publicly available at QITSA.