DualGCN: Exploring Syntactic and Semantic Information for Aspect-Based Sentiment Analysis.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Aspect-based sentiment analysis (ABSA) identifies sentiment polarities for specific aspects in text.
• Graph convolutional networks (GCNs) integrating syntactic dependency structures show promise in ABSA.
• GCN performance is sensitive to variations in dependency parsing results.

Purpose of the Study:

• To propose a DualGCN model for aspect-based sentiment analysis.
• To jointly leverage syntactic structures and semantic correlations.
• To mitigate the performance dependency on external dependency parsers.

Main Methods:

• SynGCN module implicitly integrates syntactic information using a dependency probability matrix.
• SemGCN module employs multi-head attention to enhance syntax with semantic information.
• Orthogonal and differential regularizers constrain attention scores for precise semantic correlation capture.
• Mutual BiAffine module bridges information between SynGCN and SemGCN.

Main Results:

• DualGCN demonstrates superior performance compared to state-of-the-art approaches on multiple ABSA datasets.
• Experimental results confirm the performance impact of different dependency parsers on GCN-based models.
• The proposed model achieves robust performance across various datasets, including Restaurant14, Laptop14, Twitter, Restaurant15, and Restaurant16.

Conclusions:

• The DualGCN model effectively integrates syntactic and semantic information for improved ABSA.
• The proposed approach offers a more robust solution less dependent on specific dependency parsers.
• The study provides publicly available code and datasets for reproducible research.