[Paper Review] Graph Ensemble Learning over Multiple Dependency Trees for Aspect-level Sentiment Classification

Xiaochen Hou, Peng Qi+5 more|arXiv (Cornell University)|Mar 12, 2021

Sentiment Analysis and Opinion Mining25 references4 citations

TL;DR

This paper proposes GraphMerge, a graph ensemble technique that combines dependency trees from multiple parsers into a single, unified graph before applying graph neural networks (GNNs) for aspect-level sentiment classification. By aggregating edges from diverse parsers, GraphMerge enhances robustness to parsing errors, improves model generalization, and achieves state-of-the-art performance without additional parameters or computational cost.

ABSTRACT

Recent work on aspect-level sentiment classification has demonstrated the efficacy of incorporating syntactic structures such as dependency trees with graph neural networks(GNN), but these approaches are usually vulnerable to parsing errors. To better leverage syntactic information in the face of unavoidable errors, we propose a simple yet effective graph ensemble technique, GraphMerge, to make use of the predictions from differ-ent parsers. Instead of assigning one set of model parameters to each dependency tree, we first combine the dependency relations from different parses before applying GNNs over the resulting graph. This allows GNN mod-els to be robust to parse errors at no additional computational cost, and helps avoid overparameterization and overfitting from GNN layer stacking by introducing more connectivity into the ensemble graph. Our experiments on the SemEval 2014 Task 4 and ACL 14 Twitter datasets show that our GraphMerge model not only outperforms models with single dependency tree, but also beats other ensemble mod-els without adding model parameters.

Motivation & Objective

• To address the vulnerability of dependency-based sentiment models to parsing errors in syntactic structures.
• To improve model robustness without increasing model parameters or computational cost.
• To leverage diverse parsing hypotheses from multiple parsers to enhance GNN generalization.
• To reduce overfitting and over-smoothing in GNNs by shortening graph diameter through edge ensembling.
• To outperform single-parser models and traditional ensembling methods in aspect-level sentiment classification.

Proposed method

• Construct an ensemble graph by taking the union of dependency edges from multiple parsers, preserving syntactic structure across hypotheses.
• Apply relational graph attention networks (RGAT) to the ensemble graph with separate embeddings for parent-to-child and child-to-parent edges.
• Use the unified graph as input to GNNs, allowing the model to learn to prioritize reliable edges from multiple parsing sources.
• Limit over-parameterization by avoiding separate GNNs per parse tree and reducing graph diameter through increased connectivity.
• Encode syntactic relationships using relational attention mechanisms that distinguish directional dependency relations.
• Train the model end-to-end on aspect-level sentiment classification with cross-entropy loss, using standard benchmark datasets.

Experimental results

Research questions

• RQ1 Can combining multiple dependency parses into a single graph improve robustness to parsing errors in aspect-level sentiment classification?
• RQ2 Does ensembling multiple parsing hypotheses via graph union lead to better performance than using a single parse tree?
• RQ3 Can GraphMerge achieve state-of-the-art results without adding model parameters or computational cost?
• RQ4 To what extent does GraphMerge reduce overfitting and over-smoothing in GNNs compared to stacked GNN layers?
• RQ5 How does the ensemble graph improve the model's ability to capture correct syntactic relationships, especially when individual parsers fail?

Key findings

• GraphMerge achieves state-of-the-art performance on the SemEval 2014 Restaurant and Laptop datasets, with 85.16% accuracy on 14Rest and 80.0% on 14Lap.
• On the ARTS benchmark, GraphMerge achieves an Aspect Robustness Score (ARS) of 57.46 on 14Rest and 52.90 on 14Lap, outperforming all single-parser models and BERT.
• The model reduces the average distance between aspect terms and opinion words to no more than two hops in the ensemble graph, improving message passing efficiency.
• GraphMerge improves recall of gold-standard dependency edges in challenging cases, even when no single parser predicts them correctly.
• The approach achieves better robustness than BERT and single-parse RGAT models, with a 27.70-point ARS on 14Rest and 27.10 on 14Lap, indicating strong generalization under perturbation.
• GraphMerge avoids overfitting and over-smoothing by reducing graph diameter and limiting GNN depth, requiring fewer layers than baseline models.