TY - JOUR
T1 - QDRF
T2 - Quantized and distilled robust fusion for incomplete data in multimodal sentiment analysis
AU - Hou, Jiachen
AU - Cheng, Zheyan
AU - Xing, Chen
AU - Li, Yushi
AU - Xiang, Nan
AU - Cai, Shaoyu
AU - Liu, Jie
AU - Pan, Yushan
N1 - Publisher Copyright:
© 2026 Elsevier Inc.
PY - 2027/1/5
Y1 - 2027/1/5
N2 - Despite the success of Multimodal Sentiment Analysis in ideal settings, real-world data corruption often leads to “variance explosion” in continuous latent spaces and “semantic drift” during teacher–student alignment. To overcome these challenges, we introduce Quantized and Distilled Robust Fusion (QDRF), a framework that replaces fragile continuous imputation with a Finite Scalar Quantization bottleneck to improve robustness against latent-space perturbations through a discrete semantic pivot. We further stabilize the system using an asymmetric Exponential Moving Average distillation framework that enforces dual alignment at both the feature and weight levels. Evaluated on CMU-MOSI, CMU-MOSEI, and CH-SIMS, QDRF achieves state-of-the-art or highly competitive performance on CMU-MOSEI and CH-SIMS, demonstrating that linguistic-prior-guided discrete representation learning is effective for robust sentiment prediction in incomplete data environments.
AB - Despite the success of Multimodal Sentiment Analysis in ideal settings, real-world data corruption often leads to “variance explosion” in continuous latent spaces and “semantic drift” during teacher–student alignment. To overcome these challenges, we introduce Quantized and Distilled Robust Fusion (QDRF), a framework that replaces fragile continuous imputation with a Finite Scalar Quantization bottleneck to improve robustness against latent-space perturbations through a discrete semantic pivot. We further stabilize the system using an asymmetric Exponential Moving Average distillation framework that enforces dual alignment at both the feature and weight levels. Evaluated on CMU-MOSI, CMU-MOSEI, and CH-SIMS, QDRF achieves state-of-the-art or highly competitive performance on CMU-MOSEI and CH-SIMS, demonstrating that linguistic-prior-guided discrete representation learning is effective for robust sentiment prediction in incomplete data environments.
KW - Finite scalar quantization
KW - Missing modalities
KW - Multimodal sentiment analysis
KW - Teacher-student distillation
UR - https://www.scopus.com/pages/publications/105046550840
U2 - 10.1016/j.ins.2026.123966
DO - 10.1016/j.ins.2026.123966
M3 - Article
AN - SCOPUS:105046550840
SN - 0020-0255
VL - 758
JO - Information Sciences
JF - Information Sciences
M1 - 123966
ER -