Abstract
Aspect-based Sentiment Analysis (ABSA) is a fine-grained sentiment classification task that requires capturing the relationship between specific aspects and their context. Traditional methods often rely on dependency trees to guide Graph Convolutional Networks (GCNs), but these are susceptible to parser errors and noise. We present DualGCN, a model that simultaneously leverages syntactic dependency and semantic correlation. By employing a dual-branch architecture with a BiAffine interaction module and novel regularizers, DualGCN effectively mitigates the limitations of "hard" syntactic constraints. Experimental results on three benchmark datasets demonstrate that DualGCN achieves state-of-the-art performance, particularly in challenging, informal text environments.
Introduction
The primary goal of ABSA is to determine the sentiment polarity of specific aspects within a sentence. For instance, in the review "The price is reasonable although the service is poor," a model must correctly associate "reasonable" with "price" and "poor" with "service." Current research faces two major hurdles: attention noise in complex sentences and the inherent fallibility of dependency parsers. While GCNs have improved performance by incorporating syntactic structure, they often fail when the underlying dependency tree is inaccurate or when semantic relationships transcend grammatical bounds.
Methodology: The DualGCN Framework
DualGCN introduces a parallel processing architecture designed to capture both the formal structure and the latent meaning of text.
Syntactic and Semantic Branches
- SynGCN (Syntax-based): Rather than using a binary adjacency matrix derived from a single-best dependency tree, SynGCN utilizes the probability matrix of all potential dependency arcs. This "soft" syntax approach allows the model to consider a wider range of grammatical possibilities, making it more robust to parsing errors.
- SemGCN (Semantic-based): This branch uses a self-attention mechanism to construct a semantic adjacency matrix. It identifies correlations between words that may be syntactically distant but semantically linked.
Interaction and Regularization
To harmonize these branches, the authors implement a BiAffine module that enables mutual information exchange. Furthermore, two specific regularizers are introduced to ensure the branches learn complementary features:
- Orthogonal Regularizer: Encourages sparsity in the semantic attention weights, reducing redundancy and noise.
- Differential Regularizer: Penalizes the SemGCN if its learned structure becomes too similar to the SynGCN's syntactic structure, forcing the semantic branch to capture unique contextual nuances.
Experimental Results
The model was evaluated on the Restaurant, Laptop, and Twitter datasets. DualGCN consistently outperformed previous state-of-the-art models, including ASGCN and DGEDT.
As shown in the results, DualGCN achieved an accuracy of 84.27% on the Restaurant dataset using BiLSTM and 87.13% when combined with BERT. Notably, the model showed significant improvements on the Twitter dataset, which is typically difficult for syntax-heavy models due to its informal and often ungrammatical nature.
Ablation studies confirmed that removing the BiAffine module or the regularizers led to measurable performance degradation, validating the importance of the dual-branch interaction. Furthermore, the authors observed that the model's performance peaks at two GCN layers. Increasing the depth beyond this point leads to information redundancy and the "over-smoothing" problem, where node representations become too similar to be discriminative.
Conclusion
DualGCN effectively addresses the limitations of traditional dependency-based ABSA models by integrating a probabilistic syntactic branch with a regularized semantic branch. This approach provides a more holistic and robust representation of text, leading to superior sentiment classification. Future research could explore the integration of external knowledge graphs or the application of this dual-branch strategy to other natural language understanding tasks.
Summary updated