Triple dimensional psychology knowledge encouraging graph attention networks to exploit aspect-based sentiment...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Aspect-Based Sentiment Analysis (ABSA) is a key subtask in NLP for predicting sentiment polarity of specific terms.
• Existing ABSA methods often rely on single-dimensional external affective knowledge, limiting sentiment comprehension.
• There is a need for multi-dimensional approaches to capture complex sentiment information.

Purpose of the Study:

• To propose a novel ABSA network, VADGAT, that exploits sentiment features from triple dimensions.
• To enhance sentiment representation by integrating psychology knowledge and SenticNet.
• To improve the accuracy and comprehensiveness of aspect-based sentiment analysis.

Main Methods:

• Developed an aspect-oriented template for fine-tuning pre-trained language models.
• Constructed a dependency graph with three independent adjacency matrices incorporating triple-dimensional sentiment.
• Utilized Graph Attention Networks (GAT) for relation extraction and SenticNet to enhance sentiment intensity.
• Integrated an intentional shadow network to refine sentiment information screening, optimized using Jensen-Shannon divergence.

Main Results:

• The proposed VADGAT network effectively extracts multi-dimensional sentiment features.
• SenticNet integration enhances the detection of implicit sentiment information.
• The shadow network improves the screening of relevant sentiment data.
• Extensive experiments on five datasets demonstrate superior performance over state-of-the-art methods.

Conclusions:

• The VADGAT network offers a significant advancement in aspect-based sentiment analysis by incorporating multi-dimensional sentiment knowledge.
• The approach effectively addresses the limitations of single-dimensional knowledge exploitation in previous ABSA research.
• The findings highlight the potential of integrating diverse affective knowledge sources for more robust sentiment analysis.