QiNN-QJ: A Quantum-inspired Neural Network with Quantum Jump for Multimodal Sentiment Analysis

Understanding human emotions from multiple sources like text, audio, and video is crucial for applications in mental health monitoring, customer service, and social media analysis, yet current AI systems often struggle with accuracy and transparency. Researchers have developed a quantum-inspired neural network that not only improves prediction performance but also provides interpretable results, making it a significant step forward for reliable emotion analysis in real-world scenarios.

The key finding of this study is that the proposed QiNN-QJ model achieves state-of-the-art results in multimodal sentiment analysis by simulating quantum entanglement to capture complex dependencies between different data types. Unlike traditional methods, it represents inputs as quantum-like states and uses a quantum jump mechanism to model interactions, leading to more stable and generalizable predictions. For instance, on the CMU-MOSI dataset, the model reached an accuracy of 88.2% in binary classification, outperforming other advanced systems.

Methodologically, the approach begins by encoding text, audio, and visual inputs into unit-norm complex-valued vectors, which act as quantum states. These are combined into a joint state using tensor products, and then transformed through a module that integrates Hamiltonian operators for coherent evolution and Lindblad operators for dissipative effects, mimicking open quantum system dynamics. This process allows the model to balance expressive power with stability, avoiding the overfitting common in purely unitary transformations. The final predictions are made by projecting the entangled states onto learned measurement vectors, optimized through a joint loss function that includes task-specific and contrastive components.

Results analysis shows consistent superiority across multiple benchmarks. On the CMU-MOSEI dataset, QiNN-QJ achieved a correlation coefficient of 0.811, indicating strong alignment with human-labeled sentiment intensities, and on the CH-SIMS dataset, it excelled in five-class accuracy with 46.92%, highlighting its adaptability to different languages and cultural contexts. The model's interpretability was enhanced through von Neumann entropy measurements, which quantified entanglement levels and revealed how it adapts to dataset-specific correlations, as illustrated in the paper's figures showing entropy evolution over time steps.

In practical terms, this innovation matters because it enables more trustworthy AI systems for emotion detection, which can be deployed in healthcare for mood disorder assessments or in education for personalized feedback. The model's robustness to incomplete data—maintaining performance even with up to 50% missing modalities—ensures reliability in noisy real-world environments, such as videos with poor audio quality or text with ambiguous expressions.

Limitations of the work, as noted in the paper, include the need for further exploration into variational circuit implementations and uncertainty-aware reasoning. The current framework, while effective, relies on classical hardware and may not fully exploit quantum computing advantages, leaving room for future enhancements in cross-domain applications and deeper explainability.