Abstract
Aspect-level Sentiment Classification (ASC) aims to determine the sentiment expressed toward specific aspect terms within a sentence. While dependency trees are effective at capturing long-range syntactic relations between aspects and opinion words, current models are often limited by the inaccuracies of single dependency parsers. We propose GraphMerge, a novel ensemble framework that constructs a unified graph by taking the union of edges from multiple dependency parsers. This approach leverages the "wisdom of the crowd" to mitigate parsing errors and improve structural robustness. Our results show that GraphMerge achieves state-of-the-art performance across multiple benchmarks, including the adversarial ARTS dataset, while maintaining high parameter efficiency.
Introduction: The Brittleness of Single-Parser Dependency Trees
In Aspect-level Sentiment Classification (ASC), identifying the sentiment toward a specific aspect is crucial. For instance, in the sentence "The exterior, unlike the food, is unwelcoming," the sentiment is negative for "exterior" but positive for "food." To capture these nuances, state-of-the-art models utilize dependency trees—syntactic maps that connect aspect terms directly to their relevant opinion words, even when they are separated by many tokens.
However, dependency parsers are prone to errors, particularly on informal text like tweets or online reviews. If a single parser mistakenly links an aspect to the wrong opinion word, the model's prediction will likely fail. Most existing models rely on a single parser, making them "brittle" and vulnerable to these inevitable syntactic mistakes.
Methodology: The GraphMerge Framework
To address this vulnerability, we introduce GraphMerge, a graph ensemble technique that leverages multiple syntactic perspectives from different parsers (specifically CoreNLP, Stanza, and the Berkeley neural parser). Instead of a traditional ensemble that averages model outputs, GraphMerge operates at the structural level.
The framework consists of three primary components:
- Edge Union: The model takes the union of all edges predicted by the various parsers to construct a single "ensemble graph." This ensures that if at least one parser identifies the correct syntactic relationship, the information is available to the model.
- Relational Graph Attention Network (RGAT): An RGAT is applied over this combined graph. To preserve syntactic integrity, the model distinguishes between different edge types, including parent-to-child, child-to-parent, and self-loops.
- BERT Backbone: The model utilizes BERT to generate initial word representations, which are then refined by the graph-based syntactic information.
By combining graphs before the learning phase, the GNN can learn to prioritize relevant edges and ignore the noise introduced by faulty parsers during training.
Experimental Results and Robustness
GraphMerge was evaluated on three standard benchmarks: SemEval 2014 Task 4 (Restaurant and Laptop), ACL 14 Twitter, and the Aspect Robustness Test Set (ARTS).
Key results include:
- Laptop Dataset: Achieved an accuracy of 81.35%, outperforming the best single-tree baseline by over 1.4%.
- Twitter Dataset: Reached 78.18% accuracy, demonstrating strong performance on noisy, informal text.
- Adversarial Robustness: On the ARTS dataset, GraphMerge showed a significantly lower performance drop compared to BERT-only or single-parser models, proving it is less susceptible to "distractor" aspect terms.
Key Insights: Shortening the Hop Distance
A critical finding is that GraphMerge effectively reduces the syntactic distance between aspect terms and their corresponding opinion words. By taking the union of multiple parses, there is a higher probability that at least one parser will create a "shortcut" (a 1-hop or 2-hop connection) between relevant words, facilitating more efficient information flow within the GNN.
Furthermore, GraphMerge is highly efficient. Unlike label or feature ensembles that increase model size, GraphMerge maintains the same number of parameters as a single-tree model, providing a "free" upgrade in terms of memory and parameter efficiency.
Conclusion
GraphMerge demonstrates that structural ensemble learning is a powerful and efficient way to improve the robustness of syntax-aware models. While its success depends on the diversity of the underlying parsers, the "Edge Union" approach provides a significant buffer against the errors of individual parsers. Future work may explore weighted edge schemes or apply this logic to other syntax-dependent NLP tasks like Relation Extraction.
Summary updated