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

TL;DR

RDGCN (Reinforced Dependency Graph Convolutional Network) improves Aspect-based Sentiment Analysis (ABSA) by intelligently weighting syntactic connections between words. It uses reinforcement learning to automatically find the optimal "importance decay" for words at different distances in a dependency tree, outperforming state-of-the-art models across major benchmarks.

What Problem Does This Paper Address?

Aspect-based Sentiment Analysis (ABSA) is the task of identifying the sentiment (positive, negative, or neutral) toward specific "aspects" in a sentence. For example, in the sentence "Great food but the service was dreadful," the model must correctly assign "positive" to food and "negative" to service.

While Graph Neural Networks (GNNs) have become the standard for this task by leveraging syntactic dependency trees (which link words based on grammatical relationships), they face two major hurdles:

  1. Ambiguous Weighting: Standard attention mechanisms often struggle when a sentence contains multiple aspects and conflicting opinions, sometimes "distracting" the model with irrelevant words.
  2. Sub-optimal Distance Modeling: Most models treat the distance between words in a dependency tree either as a binary (connected or not) or via simple linear/reciprocal weighting. This doesn't account for the fact that syntactic influence doesn't drop off linearly; it "blurs" as distance increases.

Methods

The authors propose RDGCN, which treats the dependency tree not just as a static map, but as a weighted graph where the weights are learned through two distinct views:

1. Distance-Importance Function (The "Reinforced" Part)

The core innovation is how the model handles the "minimum tree distance" (the shortest path between two words in the dependency tree).

  • The Criterion: The authors argue that weights should be highly discriminative (1 for distance 0, 0 for max distance) and follow a non-linear distribution where the gap between weights decreases as distance increases.
  • The Function: They design a complex function combining power-based and exponential-based sub-functions to create a "concave arc" distribution.
  • Reinforcement Learning (RL): Since the ideal "curvature" of this arc varies by dataset and task, they use a Two-Armed Bandit RL module. This module automatically searches for the optimal curvature ($K$) by observing how changes to $K$ affect validation accuracy, eliminating the need for tedious manual hyperparameter tuning.

2. Type-Importance Function

Since dependency types (e.g., "nsubj" vs. "amod") don't have a natural numerical distance, the model uses a global attention mechanism combined with a mask (to preserve the original tree topology) to learn which grammatical relationships are most salient for sentiment.

3. Feature Aggregation

These learned weights are merged into an adjacency matrix, and a standard multi-layer GCN is used to aggregate features, followed by mean pooling over the aspect terms for final classification.

Datasets

The model was evaluated on three widely used ABSA benchmarks:

  • Restaurant (SemEval 2014 Task 4)
  • Laptop (SemEval 2014 Task 4)
  • Twitter (A collection of processed tweets)

The authors used the Stanza parser to generate the initial dependency trees and GloVe (or BERT) for initial word embeddings.

Key Insights and Ideas

  • Smooth vs. Abrupt Decay: The "aha moment" is that tree-based distance importance should fall off smoothly rather than abruptly. Linear weighting (equidistant) is sub-optimal because it over-penalizes words that are slightly further away but still syntactically relevant.
  • Distance > Type: Ablation studies revealed that the distance-importance function is more critical to performance than the type-importance function. Dropping the distance component led to a ~3-4% drop in accuracy/F1, whereas dropping the type component only caused a ~1% drop.
  • RL as a Portable Tuner: By using RL to find the curvature $K$, the model becomes highly portable. The RL module converges quickly (often within 10-20 time indices), finding different optimal curvatures for different datasets (e.g., $K \approx 1.3$ for Restaurant vs. $K \approx 1.1$ for Twitter).

Key Results

RDGCN achieved state-of-the-art (SOTA) results, particularly when paired with BERT:

  • Laptop Dataset: RDGCN+BERT reached 82.12% Accuracy and 78.34% F1, significantly outperforming previous SOTA models like DualGCN and SSEGCN.
  • Twitter Dataset: It achieved 78.29% Accuracy, the highest reported in the study.
  • Case Studies: Qualitative analysis showed that RDGCN is much better at ignoring "interfering" opinion words (e.g., in "Biggest complaint is Windows 8," it correctly focuses on the negative sentiment of "complaint" rather than being neutralized by the neutral "Windows 8").

Limitations and Future Work

  • Semantic Capture: On the Restaurant dataset, the base RDGCN (without BERT) slightly lagged behind models like DualGCN. The authors attribute this to those models having better mechanisms for capturing raw semantic information via self-attention, though RDGCN+BERT eventually bridges this gap.
  • Dependency Parser Reliance: Like all syntax-based models, RDGCN is beholden to the quality of the underlying dependency parser (Stanza). Errors in the initial tree will propagate through the importance functions.
  • Future Direction: The authors suggest that exploring non-equidistant importance for other explicit syntactic features beyond just tree distance could be a fruitful path forward.

Summary updated

Abstract

Aspect-based sentiment analysis (ABSA) is dedicated to forecasting the sentiment polarity of aspect terms within sentences. Employing graph neural networks to capture structural patterns from syntactic dependency parsing has been confirmed as an effective approach for boosting ABSA. In most works, the topology of dependency trees or dependency-based attention coefficients is often loosely regarded as edges between aspects and opinions, which can result in insufficient and ambiguous syntactic utilization. To address these problems, we propose a new reinforced dependency graph convolutional network (RDGCN) that improves the importance calculation of dependencies in both distance and type views. Initially, we propose an importance calculation criterion for the minimum distances over dependency trees. Under the criterion, we design a distance-importance function that leverages reinforcement learning for weight distribution search and dissimilarity control. Since dependency types often do not have explicit syntax like tree distances, we use global attention and mask mechanisms to design type-importance functions. Finally, we merge these weights and implement feature aggregation and classification. Comprehensive experiments show the effectiveness of the criterion and importance functions. RDGCN yields excellent analysis results.
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