Convolution over Hierarchical Syntactic and Lexical Graphs for Aspect Level Sentiment Analysis

Mi Zhang, Tieyun QianEMNLP 2020 · DOI

Introduction

Aspect-Level Sentiment Classification (ASC) aims to identify the sentiment expressed toward specific entities (aspects) within a sentence. For example, in the review "the food was okay, but the service was slow," a model must distinguish between the neutral sentiment for "food" and the negative sentiment for "service."

Recent advancements have utilized dependency trees and Graph Convolutional Networks (GCNs) to capture the grammatical relationships between aspects and their descriptors. However, these approaches often struggle with two key issues: they ignore global lexical context (word collocations across the entire corpus) and they treat all syntactic relations as equal, failing to distinguish between different types of grammatical dependencies. To address these gaps, the authors propose the Bi-level interactive Graph Convolution Network (BiGCN).

The BiGCN Architecture

The BiGCN model introduces a novel way to represent sentence structure and word associations by combining local syntactic information with global lexical knowledge through a hierarchical framework.

Hierarchical Graph Construction

Instead of relying on raw, noisy dependency labels, BiGCN groups 36 specific syntactic relations into five high-level "concepts" (e.g., noun-related, verb-related). This hierarchy reduces noise and helps the model focus on the functional roles of words. Simultaneously, the model constructs a Global Lexical Graph based on word co-occurrences across the training set, capturing collocations like "nothing special" that carry specific sentiment meanings not easily derived from individual words.

Comparison of standard dependency trees with the proposed hierarchical syntactic and lexical graphs.

HierAgg Module and Bi-level GCN

The core of the model is the HierAgg (Hierarchical Aggregation) Module. It utilizes a Cross Network to fuse sequence features from a Bi-LSTM with graph-based features. This allows the syntactic structure of a specific sentence to interact with global lexical associations, refining the representation of each word.

To handle the labeled edges in these graphs, the model employs a Bi-level GCN. This mechanism first aggregates nodes sharing the same relation type into "virtual nodes" and then aggregates these virtual nodes to update the target word's representation. This two-step process ensures that the model respects the distinct nature of different grammatical and lexical relationships.

The overall framework of BiGCN and the detailed structure of the HierAgg module.

Performance and Evaluation

The BiGCN was evaluated on five benchmark datasets: Twitter, Lap14, Rest14, Rest15, and Rest16. The model achieved state-of-the-art (SOTA) performance across all datasets, with particularly significant gains in Macro-F1 scores, which measure performance across imbalanced sentiment classes.

Key findings from the evaluation include:

  • Superiority over Baselines: On the Rest16 dataset, BiGCN improved the Macro-F1 score by 3.12% compared to previous leading graph-based models.
  • Ablation Success: Removing either the hierarchical structure or the lexical graph led to a noticeable drop in accuracy, confirming that both global context and hierarchical abstraction are vital for performance.
  • Robustness: The hierarchical approach makes the model more resilient to errors in automated dependency parsing.

Conclusion

By integrating global lexical collocations with a hierarchical view of syntax, BiGCN overcomes the limitations of traditional dependency-based models in aspect-level sentiment analysis. While the model is sensitive to the number of layers in the HierAgg module—with 2 or 3 layers being optimal to avoid overfitting—it represents a significant step forward in making graph-based NLP models more nuanced and context-aware.

Summary updated

BibTeX
@inproceedings{zhang2020convolution,
  author = {Mi Zhang and Tieyun Qian},
  title = {{Convolution over Hierarchical Syntactic and Lexical Graphs for Aspect Level Sentiment Analysis}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.emnlp-main.286},
}
Figure