International Conference on Mathematics, Actuarial Science, Computer Science and Statistics (MACS)
Quantum Infused Multi-Modal Sentiment Analysis of Low Resourced Language
Muzammal Hussain, Imran Khan, Farman Shah, Rana Mudassar Rasool, Ahmad Khurram, Saad Rehman
Abstract
For Urdu sentiment analysis, we present a quantum-enhanced multimodal system that combines a variational quantum circuit (VQC) classifier with traditional feature extractors. To create a cohesive embedding, a cross-modal transformer model first processes raw text, audio, and visual information. On a noisy intermediate-scale quantum (NISQ) processor, these fused embeddings are then angle-encoded into a parameterized quantum state. Simultaneous perturbation stochastic approximation (SPSA) is used to improve the trainable rotation gates of the VQC under realistic hardware-inspired noise models. Our quantum-enhanced model outperforms the classical cross-modal baseline$\mathbf{ (8 2. 4 \%} \boldsymbol{/} \mathbf{7 9. 6 \%} \boldsymbol{)}$and a purely classical MLP head (83.1%/80.8%) with an overall accuracy of 84.2% and a macro-F1 score of 81.7% when tested on UMSA-U, a curated dataset of 6,000 Urdu video clips with aligned transcripts, audio, and facial expression annotations. The findings show that noise-robust SPSA tuning maintains performance on simulated hardware error rates up to 5%, while quantum feature encoding produces more separable decision boundaries. These findings establish a new standard for Urdu sentiment recognition on NISQ devices and demonstrate the potential of variational quantum classifiers for low-resource, multimodal language solutions.