Exploring all-quantum long short-term memory networks to enhance text sentiment classification

Quantum natural language processing (QNLP) can efficiently represent and process in parallel a vast number of words and sentences in high-dimensional Hilbert space by leveraging the characteristics of quantum superposition and entanglement, thereby accelerating complex text sentiment classification tasks. This study, for the first time, proposes a full quantum model based on long short-term memory network for text sentiment classification, where the fundamental representation learning and sequence modeling operations are all realized through quantum state evolution. Additionally, a subspace projective quantum self-attention mechanism is designed to effectively capture long-distance dependencies and semantic associations in text sequences. This method achieves higher accuracy, reduced circuit complexity, and expanded scalability with larger datasets. Extensive experiments demonstrate that the proposed approach outperforms existing quantum-enhanced models in terms of overall performance and exhibits strong robustness against various quantum noise interferences. Our model is anticipated to be effectively implemented on noisy intermediate-scale quantum (NISQ) devices to yield results of practical significance, thereby advancing the development of the QNLP field.