C2MF: Confidence-Calibrated Multimodal Fusion for Continuous Sentiment Regression

View at publisher Summarise with VIRAT
Authors: Jiaxiong Liu, Ke Qi, Zhiwen Liao, Feixiang Yuan, Wen Zhuo, Yueyue Zhou
Cited by: 0
References: 38

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

To alleviate noise amplification and modality imbalance in continuous multimodal sentiment regression, we propose C2MF, a confidence-calibrated fusion framework that explicitly models sample-wise modality reliability. C2MF introduces aContinuous Modality-Confidence Gatingmodule (CMCG) that derives visual/acoustic confidence scores from global descriptors and injects them into cross-modal attention to suppress unreliable streams while retaining informative cues. To further stabilize regression, we adoptlogit-level mixupto regularize the output space and enforce local smoothness of sentiment scores, together with apolarity-aware joint objectivethat couples coarse polarity supervision with continuous intensity learning. Experiments on CMU-MOSI and CMU-MOSEI validate the effectiveness of C2MF, yielding consistent improvements over strong baselines in MAE, correlation, and accuracy metrics (Acc2/Acc7).

Fields of Study & Subjects

Sentiment Analysis and Opinion Mining Emotion and Mood Recognition Multimodal Machine Learning Applications

Publication details

Journal / source
IEEE Signal Processing Letters
Volume
33
Pages
1356–1360
Publication date
1 January 2026
Language
EN