A Quantum-Inspired Residual Self-Attention Network for Multimodal Sentiment Analysis

Multimodal sentiment analysis is challenging because textual, visual, and acoustic evidence is heterogeneous and oft en weakly aligned. Here, we present QRSAN, a quantum-inspired residual self-attention network that integrates an LSTM text encoder, modality-specific multilayer perceptrons for visual and acoustic inputs, normalized complex-valued encoding, a TFN-derived interaction expansion, real-valued residual self-attention, and classical projection-score mapping. QRSAN runs entirely on classical hardware and does not perform physical quantum computation. Across CMU-MOSI, CMU-MOSEI, and IEMOCAP, QRSAN was evaluated using a common protocol. It achieved the highest numerical mean ACC and Binary_F1 among the evaluated models on CMU-MOSI, whereas its IEMOCAP label-wise accuracy was below that of EF-LSTM. These findings support the utility of combining constrained complex-valued representations with residual self-attention using the evaluated settings, without claiming universal state-of-the-art performance.