Scholarlink

<p>Aspect-based Sentiment Analysis (ABSA) focuses on identifying the sentiment polarity associated with specific aspect terms in user reviews; this fine-grained sentiment analysis is crucial for product optimization. Although existing models have achieved promising performance, they still encounter challenges in effectively addressing cross-aspect sentiment interference when multiple aspects coexist, as well as in adequately fusing multi-source features. This study proposes a dual-enhanced multi-view graph neural network model, which substantially enhances the alignment capability between aspect terms and sentiment expressions, thereby enabling more accurate fine-grained semantic extraction. Specifically, to obtain high-quality task-oriented initial semantic representations, a dual-enhanced encoding module integrating noise resilience and sentiment awareness. Moreover, to improve the alignment between aspect terms and sentiments, a multi-view graph structure modeling module is proposed. The global semantic branch utilizes multi-head attention to capture long-range dependencies. The local semantic branch achieves precise alignment between aspects and their contextual expressions through dynamic boundary learning and phrase structure tree supervision. The syntactic branch integrates dependency relations with semantic attention, introducing a Laplace variant with root node preference, marginal probability modeling, and root node constraints to dynamically optimize a more aspect-aware syntactic graph structure. Finally, a multi-view gated and global-aware collaborative fusion module is proposed, which dynamically aggregates multi-perspective features and enhances global perception capabilities. Experimental results on four public datasets demonstrate that our method outperforms existing baselines across multiple evaluation metrics. In addition, to efficiently handle large-scale textual datasets and enable real-time sentiment analysis, our framework benefits from high-performance computing (HPC) and parallel processing.</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

DMGM: dual-enhanced multi-view graph modeling for aspect-based sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Wei Zheng,
  • Chuanfei Liu,
  • Zhenlin Zhang

摘要

Aspect-based Sentiment Analysis (ABSA) focuses on identifying the sentiment polarity associated with specific aspect terms in user reviews; this fine-grained sentiment analysis is crucial for product optimization. Although existing models have achieved promising performance, they still encounter challenges in effectively addressing cross-aspect sentiment interference when multiple aspects coexist, as well as in adequately fusing multi-source features. This study proposes a dual-enhanced multi-view graph neural network model, which substantially enhances the alignment capability between aspect terms and sentiment expressions, thereby enabling more accurate fine-grained semantic extraction. Specifically, to obtain high-quality task-oriented initial semantic representations, a dual-enhanced encoding module integrating noise resilience and sentiment awareness. Moreover, to improve the alignment between aspect terms and sentiments, a multi-view graph structure modeling module is proposed. The global semantic branch utilizes multi-head attention to capture long-range dependencies. The local semantic branch achieves precise alignment between aspects and their contextual expressions through dynamic boundary learning and phrase structure tree supervision. The syntactic branch integrates dependency relations with semantic attention, introducing a Laplace variant with root node preference, marginal probability modeling, and root node constraints to dynamically optimize a more aspect-aware syntactic graph structure. Finally, a multi-view gated and global-aware collaborative fusion module is proposed, which dynamically aggregates multi-perspective features and enhances global perception capabilities. Experimental results on four public datasets demonstrate that our method outperforms existing baselines across multiple evaluation metrics. In addition, to efficiently handle large-scale textual datasets and enable real-time sentiment analysis, our framework benefits from high-performance computing (HPC) and parallel processing.