Scholarlink

<p>The increasing availability of multilingual and multimodal data on online platforms has created a need for the development of sophisticated models for cross-lingual sentiment analysis. Although existing sentiment analysis models are largely dependent on monolingual text data, they tend to fail in generalizing across languages and modalities, especially in low-resource and code-switching settings. This work aims to handle the most important task of combining heterogeneous modalities of text and audio data in multiple languages for accurate and interpretable sentiment classification. X-Sent-Fuse, a cross-lingual sentiment fusion model using Transformers, combines text and audio modalities to make accurate, robust, and culturally informed sentiment predictions for a wide range of languages. The architecture incorporates a Multimodal Transformer Cross-Attentive Gated Fusion (MT-CAG2F) module to effectively align semantic and prosodic features, followed by a BiSRO-Net classifier, to ensure stable and high-performing predictions. A lightweight ensemble of multilingual large language models further improves linguistic adaptability. Extensive evaluations on four benchmark datasets, Sentiment analysis in Twitter, Amazon Reviews, mTEDx, and CMU-MOSEAS, demonstrate state-of-the-art performance, achieving up to 98.67% overall accuracy and 0.996 AUC, with consistently high precision, recall, and F1-scores. These results validate the effectiveness of X-Sent-Fuse in handling diverse and noisy cross-lingual inputs, offering a scalable solution for real-world applications such as multilingual customer feedback analysis, cross-cultural media monitoring, and affective computing in global communication platforms.</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

Cross-lingual sentiment analysis via multimodal transformer fusion and lightweight deep ensemble learning framework

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Rashmi Yadav,
  • Deepak Kumar Yadav,
  • Ati Jain,
  • Lalji Prasad

摘要

The increasing availability of multilingual and multimodal data on online platforms has created a need for the development of sophisticated models for cross-lingual sentiment analysis. Although existing sentiment analysis models are largely dependent on monolingual text data, they tend to fail in generalizing across languages and modalities, especially in low-resource and code-switching settings. This work aims to handle the most important task of combining heterogeneous modalities of text and audio data in multiple languages for accurate and interpretable sentiment classification. X-Sent-Fuse, a cross-lingual sentiment fusion model using Transformers, combines text and audio modalities to make accurate, robust, and culturally informed sentiment predictions for a wide range of languages. The architecture incorporates a Multimodal Transformer Cross-Attentive Gated Fusion (MT-CAG2F) module to effectively align semantic and prosodic features, followed by a BiSRO-Net classifier, to ensure stable and high-performing predictions. A lightweight ensemble of multilingual large language models further improves linguistic adaptability. Extensive evaluations on four benchmark datasets, Sentiment analysis in Twitter, Amazon Reviews, mTEDx, and CMU-MOSEAS, demonstrate state-of-the-art performance, achieving up to 98.67% overall accuracy and 0.996 AUC, with consistently high precision, recall, and F1-scores. These results validate the effectiveness of X-Sent-Fuse in handling diverse and noisy cross-lingual inputs, offering a scalable solution for real-world applications such as multilingual customer feedback analysis, cross-cultural media monitoring, and affective computing in global communication platforms.