Scholarlink

<p>Aspect-level sentiment classification is a fine-grained sentiment analysis task that aims at judging the sentiment polarity of aspects in a given sentence. Recent studies are more focused on modeling the syntax information of the sentences based on dependency trees by leveraging graph neural networks. However, these methods still have the following limitations: (1) When aspect terms contain more than one word, the normal encoding approach not only introduces noise such as prepositions, but also results in the loss of key information, which hinders the accurate representation learning of the sentences. (2) Previous methods that model the syntax information of the sentences fail to consider the presence of other linguistic features, which leads to sub-optimal prediction results. To address the restrictions mentioned earlier, we propose a graph-based method, called Co-MFGCN. More specifically, to deal with the limitation 1, we adopt co-attention mechanism to allow full word-level information interaction between aspect terms and contexts at multiple levels. To handle the limitation 2, we propose a new approach to fully exploit linguistic features to enhance our graph-based model, including part-of-speech combination, syntactic dependency types, and tree-based distance feature information. A large number of experiments are conducted on five datasets, and the results verify that Co-MFGCN outperforms multiple SOTA baseline models.</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

Co-MFGCN: multi-feature channel graph convolutional networks based on co-attention for aspect-level sentiment classification

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Yuling Zhang,
  • Fang’ai Liu

摘要

Aspect-level sentiment classification is a fine-grained sentiment analysis task that aims at judging the sentiment polarity of aspects in a given sentence. Recent studies are more focused on modeling the syntax information of the sentences based on dependency trees by leveraging graph neural networks. However, these methods still have the following limitations: (1) When aspect terms contain more than one word, the normal encoding approach not only introduces noise such as prepositions, but also results in the loss of key information, which hinders the accurate representation learning of the sentences. (2) Previous methods that model the syntax information of the sentences fail to consider the presence of other linguistic features, which leads to sub-optimal prediction results. To address the restrictions mentioned earlier, we propose a graph-based method, called Co-MFGCN. More specifically, to deal with the limitation 1, we adopt co-attention mechanism to allow full word-level information interaction between aspect terms and contexts at multiple levels. To handle the limitation 2, we propose a new approach to fully exploit linguistic features to enhance our graph-based model, including part-of-speech combination, syntactic dependency types, and tree-based distance feature information. A large number of experiments are conducted on five datasets, and the results verify that Co-MFGCN outperforms multiple SOTA baseline models.