软件导刊2026，Vol.25Issue(8)：7-14,8.DOI:10.11907/rjdk.261176
面向多模态情感分析的元学习单模态标签生成
Meta-Learning for Unimodal Label Generation in Multimodal Sentiment Analysis
摘要
Abstract
Addressing the issue that multi-modal sentiment analysis datasets only provide fused labels and lack single-modal artificial la-bels,making it difficult to train multi-task frameworks,this paper propose a meta-learning single-modal label generation method,MLCG.This method narrows the gap between single-modal and multi-modal representations through contrastive projection,and utilizes meta-labels to correct the network to generate more reliable single-modal supervision under the constraints of bi-level optimization and denoising tasks.Ex-perimental results show that MLCG achieves a Mean Absolute Error(MAE)of 0.703,a Correlation coefficient(Corr)of 0.827,and an Accu-racy-2(Acc-2)of 86.79％on the CMU-MOSI dataset;an MAE of 0.525 and an Acc-2 of 86.05％on the CMU-MOSEI dataset;and an MAE of 0.401 and an Acc-2 of 82.10％,with an F1 score of 81.95％on the CH-SIMS dataset.The generated labels reduce the MAE to 0.128,0.208,and 0.185 on text,audio,and visual modalities,respectively,demonstrating stable performance in Chinese fine-grained scenarios.This method can improve prediction performance and label quality in the absence of single-modal artificial labels.关键词
多模态情感分析/单模态标签生成/元学习/双层优化/对比学习Key words
multimodal sentiment analysis/unimodal label generation/meta-learning/bi-level optimization/contrastive learning分类
信息技术与安全科学引用本文复制引用
曾碧卿,王嘉桢,罗智康,陈威海..面向多模态情感分析的元学习单模态标签生成[J].软件导刊,2026,25(8):7-14,8.基金项目
广东省科技计划项目(2025A0505020079) （2025A0505020079）
