Scholarlink

<p>Multimodal Sentiment Analysis (MSA) still faces three major challenges in existing research: insufficient utilization of lexical-level sentiment knowledge, lack of flexibility in cross-modal fusion, and inconsistency between prediction targets and evaluation metrics. These issues limit the generalization ability of models in complex scenarios. To address these problems, this paper proposes LUCID, a unified framework that integrates <i>Lexicon-aware Multi-tag Embedding (LME)</i>, <i>Layer-Selective Cross-modal Fusion (LSCF)</i>, and <i>Distributional Ordinal Regression with Correlation and Sign-Consistency Margin (DORC-SCM)</i>. The framework strengthens textual representations with lexicon priors, dynamically injects visual and acoustic features across layers, and jointly optimizes ordinal regression with correlation and sign-consistency constraints, thereby achieving robust multimodal modeling. Experiments on CMU-MOSI, CMU-MOSEI, and the Chinese dataset CH-SIMS demonstrate that LUCID outperforms most existing methods on key metrics such as ACC2, ACC7, MAE, and Corr. Further ablation studies and visualization analyses confirm the independent contributions and synergistic effects of each module, indicating that the framework achieves higher consistency and stability in cross-lingual multimodal sentiment analysis and shows broad application potential.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LUCID: lexicon-augmented fusion with ordinal calibration for multimodal sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Feixiang Yuan,
  • Ke Qi,
  • Jiaxiong Liu,
  • Changkai Lu,
  • Wenbin Chen

摘要

Multimodal Sentiment Analysis (MSA) still faces three major challenges in existing research: insufficient utilization of lexical-level sentiment knowledge, lack of flexibility in cross-modal fusion, and inconsistency between prediction targets and evaluation metrics. These issues limit the generalization ability of models in complex scenarios. To address these problems, this paper proposes LUCID, a unified framework that integrates Lexicon-aware Multi-tag Embedding (LME), Layer-Selective Cross-modal Fusion (LSCF), and Distributional Ordinal Regression with Correlation and Sign-Consistency Margin (DORC-SCM). The framework strengthens textual representations with lexicon priors, dynamically injects visual and acoustic features across layers, and jointly optimizes ordinal regression with correlation and sign-consistency constraints, thereby achieving robust multimodal modeling. Experiments on CMU-MOSI, CMU-MOSEI, and the Chinese dataset CH-SIMS demonstrate that LUCID outperforms most existing methods on key metrics such as ACC2, ACC7, MAE, and Corr. Further ablation studies and visualization analyses confirm the independent contributions and synergistic effects of each module, indicating that the framework achieves higher consistency and stability in cross-lingual multimodal sentiment analysis and shows broad application potential.