Graph convolutional network based on self-attention variational autoencoder and capsule contrastive learning for aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) predicts sentiment polarity by aligning aspect words with their matching sentiment words in sentences at a fine-grained level. Previous methods have limitations, such as over-reliance on syntactic trees ignoring semantic and global information, and difficulty i...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Expert systems with applications Jg. 279; S. 127172
Hauptverfasser: Wang, Xinyue, Liu, Long, Chen, Zhuo, Wang, Haiyan, Yu, Bin
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 15.06.2025
Schlagworte:
Aspect-based sentiment analysis
Capsule contrastive learning
Graph convolutional network
Variational autoencoder
Variational autoencoder
Graph convolutional network
Capsule contrastive learning
Aspect-based sentiment analysis
ISSN:0957-4174
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Aspect-based sentiment analysis (ABSA) predicts sentiment polarity by aligning aspect words with their matching sentiment words in sentences at a fine-grained level. Previous methods have limitations, such as over-reliance on syntactic trees ignoring semantic and global information, and difficulty in solving long-distance dependency problems. In response to these issues, this article puts forward a hybrid graph convolutional network (GCN) called SVCCL-GCN, which combines an improved variational autoencoder (VAE) and a contrastive learning module. First, construct dependency graph structures using syntactic dependency trees to obtain explicit syntactic dependency information. Then, a VAE module based on the self-attention mechanism is devised to acquire both global and local information of the data, and it is beneficial for capturing the explicit associations between aspect terms and relevant contexts to solve the long-distance dependency problems. Finally, so as to stimulate the model's capacity to flexibly learn comprehensive semantic information and emotional feature information between different words, a contrastive learning loss based on the capsule network is constructed by taking the output of the capsule network as the contrastive feature for unsupervised contrastive learning. The extensive experimental findings on five benchmark datasets demonstrate that SVCCL-GCN productively improves the accuracy in distinguishing sentiment polarity. The source codes and datasets are available at https://github.com/YuBinLab-QUST/SVCCL-GCN/.
ISSN:0957-4174
DOI:10.1016/j.eswa.2025.127172