DAGCN: Distance-based and Aspect-oriented Graph Convolutional Network for Aspect-based Sentiment Analysis

Zhihao Wang, Bo Zhang, Ru Yang, Chang Guo, Maozhen LiNAACL 2024 · DOI

TL;DR

This paper introduces DAGCN, a novel graph convolutional network for aspect-based sentiment analysis (ABSA). DAGCN effectively identifies opinion words relevant to specific aspects by leveraging both syntactic distance and attention mechanisms, outperforming existing methods on benchmark datasets.

What Problem Does This Paper Address?

Aspect-based sentiment analysis (ABSA) aims to determine the sentiment (positive, neutral, negative) towards specific aspects within a text. A key challenge in ABSA is accurately identifying the opinion words that express sentiment towards a given aspect. Traditional methods often struggle with noise from irrelevant words or fail to capture the nuanced relationship between aspects and opinion words, especially when they are not syntactically close. Existing graph convolutional network (GCN) approaches, while leveraging dependency trees, can be hindered by the inherent noise in these structures.

Methods

The authors propose DAGCN, a model that combines two main components:

  • Distance-based Syntactic Weight (DSW): This novel approach leverages the syntactic distance between words in a dependency tree. It assigns higher weights to words that are closer to the aspect, assuming they are more likely to be opinion words. This helps to prune irrelevant words and focus on the local context.
  • Aspect-Fusion Attention (AF): To address cases where opinion words are distant from the aspect, AF is introduced. It uses an attention mechanism to discern candidate opinion words within both local and long-range scopes, ensuring that distant but relevant opinion words are not overlooked.

These two components are used to construct an adjacency matrix, which is then fed into a Aspect-oriented Graph Convolutional Network (AoGCN). Additionally, a Self-attention GCN (SaGCN) is employed to capture semantic features, especially useful for texts with less structured syntax. A BiAffine Module is used to effectively fuse the information from AoGCN and SaGCN. A KL divergence loss is also incorporated to ensure that AoGCN and SaGCN learn distinct features.

Datasets

The authors evaluate their model on three publicly available benchmark datasets:

  • Restaurant reviews
  • Laptop reviews
  • Twitter sentiment analysis

Key Insights and Ideas

  • Syntactic distance is a strong indicator of opinion words: The paper's statistical analysis shows that opinion words are often syntactically close to the aspects they modify, making syntactic distance a valuable feature.
  • Pruning dependency trees is crucial: Dependency trees, while useful, contain noise. The DSW mechanism effectively prunes irrelevant edges, leading to more precise opinion word identification.
  • Balancing local and global context is key: The combination of DSW (local focus) and AF (handling distant opinions) allows DAGCN to capture sentiment effectively across different scopes.
  • Separate syntactic and semantic learning is beneficial: Using two distinct GCNs (AoGCN for syntax, SaGCN for semantics) and a BiAffine module to fuse their outputs leads to a more robust model.

Key Results

DAGCN achieves state-of-the-art or competitive results on the Restaurant, Laptop, and Twitter datasets. For instance, on the Restaurant dataset, DAGCN achieves an accuracy of 84.72% and a Macro-F1 score of 78.08%. Ablation studies demonstrate that each component (DSW, AF, KL divergence loss) contributes significantly to the model's performance.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Performance on informal text: DAGCN's performance on the Twitter dataset with a BERT encoder is slightly lower than APARN, potentially due to CoreNLP's limitations in parsing informal language. Future work aims to leverage AMR for better handling of such texts.
  • Hyperparameter tuning: The model has several hyperparameters (α, β, γ) that require extensive tuning. Future work will explore adaptive parameter tuning.
  • Scope of application: DAGCN currently focuses on sentiment classification. Future work plans to adapt it for end-to-end ABSA and Aspect Term Extraction (ASTE) tasks.

Summary updated

BibTeX
@inproceedings{wang2024dagcn,
  author = {Zhihao Wang and Bo Zhang and Ru Yang and Chang Guo and Maozhen Li},
  title = {{DAGCN: Distance-based and Aspect-oriented Graph Convolutional Network for Aspect-based Sentiment Analysis}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-naacl.120},
}
Figure