RDGCN: Reinforced Dependency Graph Convolutional Network for Aspect-based Sentiment Analysis

Xusheng Zhao, Hao Peng, Qiong Dai, Xu Bai, Huailiang Peng, Yanbing Liu, Qinglang Guo, Philip S. Yu

Introduction

Aspect-based Sentiment Analysis (ABSA) is a fine-grained sentiment analysis task that aims to identify the sentiment polarity (e.g., positive, negative, or neutral) associated with specific aspect terms in a sentence. For instance, in the review "Great food but the service was dreadful," a model must distinguish the positive sentiment toward food from the negative sentiment toward service.

While Graph Neural Networks (GNNs) leveraging syntactic dependency trees have become a standard approach, they often struggle with two issues: ambiguous weighting of syntactic connections and sub-optimal modeling of word distances within the tree. Standard models often fail to distinguish between relevant and irrelevant syntactic neighbors, especially in complex sentences with multiple conflicting opinions.

The RDGCN Framework

To address these challenges, the Reinforced Dependency Graph Convolutional Network (RDGCN) introduces a method to intelligently weight syntactic connections. Unlike static graph models, RDGCN treats the dependency tree as a weighted graph where the importance of connections is learned through both distance-based and type-based functions.

The overall architecture of the RDGCN model, illustrating the flow from sentence encoding to the reinforced weight adjustment and final classification.

Distance-Importance via Reinforcement Learning

The primary innovation of RDGCN is the Distance-Importance Function. The authors posit that the syntactic influence of a word should decay non-linearly as its distance from the aspect term in the dependency tree increases. They propose a "concave arc" distribution for these weights, ensuring that the model remains discriminative while allowing for a smooth transition of influence.

Comparison of various weight decay functions. The RDGCN utilizes a non-linear distribution to model how syntactic importance 'blurs' as tree distance increases.

Because the optimal "curvature" ($K$) of this weight decay varies depending on the specific linguistic characteristics of a dataset, RDGCN employs a Two-Armed Bandit Reinforcement Learning (RL) module. This module automatically searches for the optimal $K$ value by monitoring validation accuracy, effectively automating hyperparameter tuning and making the model more adaptable across different domains.

Type-Importance and Feature Aggregation

In addition to distance, the model considers dependency types (e.g., "nsubj", "amod"). Since these are categorical, a global attention mechanism with a topological mask is used to learn the relative importance of different grammatical relationships. These learned weights are integrated into an adjacency matrix, which guides a multi-layer GCN in aggregating features for final sentiment classification.

Performance and Insights

RDGCN was evaluated on the Restaurant, Laptop, and Twitter benchmarks. When integrated with BERT embeddings, the model achieved state-of-the-art results:

  • Laptop: 82.12% Accuracy and 78.34% F1.
  • Twitter: 78.29% Accuracy.

Ablation studies confirmed that the distance-importance component is the most critical factor, contributing significantly more to the model's performance than type-importance. Furthermore, the RL module demonstrated rapid convergence, finding distinct optimal curvatures for different datasets (e.g., $K \approx 1.3$ for Restaurant vs. $K \approx 1.1$ for Laptop), highlighting the necessity of dataset-specific tuning.

Conclusion

RDGCN advances ABSA by moving beyond static syntactic trees. By using reinforcement learning to tune the "blurring" of syntactic influence over distance, the model achieves a more nuanced understanding of sentence structure, effectively filtering out irrelevant "noise" and focusing on the most salient opinion words for a given aspect.

Summary updated

BibTeX
@inproceedings{zhao2024rdgcn,
  author = {Xusheng Zhao and Hao Peng and Qiong Dai and Xu Bai and Huailiang Peng and Yanbing Liu and Qinglang Guo and Philip S. Yu},
  title = {{RDGCN: Reinforced Dependency Graph Convolutional Network for Aspect-based Sentiment Analysis}},
  booktitle = {WSDM},
  year = {2024},
  doi = {10.1145/3616855.3635775},
}
Figure