Scholarlink

<p>This paper presents a graph-based deep learning framework for multi-domain sentiment analysis that integrates manifold learning and capsule aggregation within a unified architecture. The model captures contextual semantics using Bidirectional Encoder Representations from Transformers (BERT) embeddings, constructs word co-occurrence graphs to encode relational dependencies, and applies graph convolutional networks (GCNs) for feature propagation. A manifold learning module preserves the topological structure during dimensionality reduction, while a capsule network with long short-term memory (LSTM)-based dynamic routing aggregates hierarchical and sequential features. Experiments conducted on two benchmark datasets—Multimodal EmotionLines Dataset (MELD) and Amazon Reviews—demonstrate that the proposed method consistently outperforms existing baselines across multiple evaluation metrics, including accuracy, precision, recall, and F1-score. Despite incorporating computationally intensive components, the model maintains competitive training and inference efficiency. The proposed approach offers a scalable, domain-resilient, and structurally aware solution for sentiment classification across diverse textual environments.</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

Manifold-based graph convolutional network for multi-domain sentiment analysis

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Sanaz Gouran Shourakchali,
  • Kamran Layeghi,
  • Faraein Aeini

摘要

This paper presents a graph-based deep learning framework for multi-domain sentiment analysis that integrates manifold learning and capsule aggregation within a unified architecture. The model captures contextual semantics using Bidirectional Encoder Representations from Transformers (BERT) embeddings, constructs word co-occurrence graphs to encode relational dependencies, and applies graph convolutional networks (GCNs) for feature propagation. A manifold learning module preserves the topological structure during dimensionality reduction, while a capsule network with long short-term memory (LSTM)-based dynamic routing aggregates hierarchical and sequential features. Experiments conducted on two benchmark datasets—Multimodal EmotionLines Dataset (MELD) and Amazon Reviews—demonstrate that the proposed method consistently outperforms existing baselines across multiple evaluation metrics, including accuracy, precision, recall, and F1-score. Despite incorporating computationally intensive components, the model maintains competitive training and inference efficiency. The proposed approach offers a scalable, domain-resilient, and structurally aware solution for sentiment classification across diverse textual environments.