Dual-Tree-Guided Aspect-Based Sentiment Analysis Incorporating Structure-Aware Semantic Refinement and Graph Attention

Xinyu Wang, Yuhang Gao, Xiao Lv, Kun Zhong, Peng Hui Tan, Zhaobin TanSymmetry 2025 · DOI

TL;DR

This paper introduces DySynGAT, a novel framework for Aspect-Based Sentiment Analysis (ABSA) that tackles the challenge of accurately linking opinions to aspects, especially in complex sentences with long-distance relationships. By synergistically combining information from both dependency and constituent syntactic trees, DySynGAT significantly improves sentiment prediction accuracy, outperforming existing state-of-the-art models.

What Problem Does This Paper Address?

Aspect-Based Sentiment Analysis (ABSA) aims to determine the sentiment (positive, negative, neutral) expressed towards specific aspects within a text. While recent advancements using pre-trained language models like BERT have improved contextual understanding, and graph neural networks (GNNs) have helped model aspect-opinion relationships, significant challenges remain. One major issue is the "broken symmetry" between syntactic and semantic representations.

Dependency trees, while good at capturing fine-grained grammatical relations, can be noisy, leading to incorrect aspect-opinion associations, especially when they are far apart in a sentence. Constituent trees, on the other hand, offer a hierarchical structure that can better define phrase boundaries but can lead to fragmentation of semantic information. This paper specifically addresses the problem of accurately capturing long-distance aspect-opinion pairs and resolving ambiguities arising from noisy or fragmented syntactic structures.

Methods

DySynGAT is built upon a multi-component architecture designed to fuse semantic and syntactic information effectively:

  • Input Encoder: A pre-trained BERT model is used to generate contextualized word embeddings.
  • Dynamic Semantic Enhancement: This module aims to repair "fragmented constituents" by:
  • Bidirectional Timing Dependency Modeling: Using BiLSTM to capture sequential context.
  • Channel Recalibration (SEN): Dynamically adjusting feature channel weights to suppress noise and amplify informative features.
  • Multi-Scale Trend Sensing Attention (MTA): Employing multi-scale convolutions and attention to capture semantic patterns at different granularities.
  • Syntax-Aware Graph Aggregation: This is a core innovation. It fuses information from both dependency and constituent trees using Localized Graph Attention Networks (LGAT). LGAT prioritizes syntactically adjacent nodes over global dependencies to mitigate over-smoothing issues common in standard GATs. It dynamically aggregates multi-level syntactic information, aligning syntactic structures with semantic representations.
  • Cross-Aspect Relationship Modeling: This module handles sentences with multiple aspects:
  • Phrase Segmentation (MSS): Identifies minimal semantic units connecting aspects.
  • Aspect–Context Interaction Graph (ACIG): Constructs a dynamic sparse graph to model interactions between aspects and their context.
  • Spatial Reduction Attention (SRA): Compresses sequential features and reduces noise in long-range dependencies.
  • LGAT Secondary Aggregation: A second pass of LGAT is used to aggregate cross-clause dependencies.
  • Classification of Emotional Polarity: Finally, the enhanced semantic and syntactic representations are fused, and a softmax classifier predicts the sentiment polarity.

Datasets

The model was evaluated on three publicly available English datasets:

  • Restaurant (SemEval2014 Task4)
  • Laptop (SemEval2014 Task4)
  • Twitter
  • MAMS

Key Insights and Ideas

  • Dual-Tree Fusion is Crucial: The paper strongly advocates for the synergistic use of both dependency and constituent trees. Neither tree type alone is sufficient; their combined strengths are needed to overcome individual limitations (noise in dependency, fragmentation in constituent).
  • Localized Attention for Noise Reduction: LGAT's focus on local syntactic neighborhoods, rather than just global dependencies, is key to mitigating noise from incorrect parsing or irrelevant cross-clause connections.
  • Dynamic Semantic Enhancement Matters: The combination of BiLSTM, SEN, and MTA effectively captures rich, context-aware, and dynamically weighted semantic representations, crucial for understanding nuanced sentiment.
  • Handling Multi-Aspect Sentences: The ACIG and SRA modules are vital for disentangling complex interactions between multiple aspects and their associated opinions, especially in sentences with contrasting clauses.

Key Results

DySynGAT demonstrated superior performance across the tested datasets. Specifically, it achieved F1 score improvements of:

  • 3.55% over SAGAT-BERT on the Laptop dataset.
  • 0.55% over SAGAT-BERT on the Restaurant dataset.
  • 1.75% over SAGAT-BERT on the Twitter dataset.

The paper also highlights that DySynGAT consistently outperforms other baseline models, including DualGCN and TD-GAT, underscoring the effectiveness of its dual-tree fusion and localized attention mechanisms. Ablation studies confirmed the significant contribution of each component, with the semantic enhancement module and LGAT being particularly critical.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while LGAT mitigates over-smoothing, the performance can still be affected by the accuracy of the initial syntactic parsers.
  • The case study shows that variants without semantic modules struggle with contrastive meanings and negations, highlighting the importance of robust semantic understanding.
  • Future Work:
    • The authors aim to improve the model's ability to process longer texts by enhancing the input encoder to better capture semantic information in extended contexts.

Summary updated

Abstract

This work addresses the broken symmetry in syntactic–semantic representations for Aspect-Based Sentiment Analysis, where advancements have been driven by the use of pre-trained language models to achieve contextual understanding and graph neural networks capturing aspect–opinion dependencies using syntactic trees. However, long-distance aspect–opinion pairs pose challenges: the structural noise in dependency trees often causes erroneous associations, while the discrete structure of the constituent trees leads to constituent fragmentation. In this paper, we propose DySynGAT and introduce a Localized Graph Attention Network (LGAT) to fuse bi-gram syntactic and semantic information from both dependency and constituent trees, effectively mitigating interference from dependency tree noise. A dynamic semantic enhancement module efficiently integrates local and global semantics, alleviating constituent fragmentation caused by constituent trees. An aspect–context interaction graph (ACIG), built upon minimal semantic segmentation and jointly enhanced features, filters out noisy cross-clause edges. Spatial reduction attention (SRA) with mean pooling compresses the redundant sequential features, reducing the noise under long-range dependencies. Experiments on foods and beverages, electronics, and user review datasets demonstrate F1 score improvements of 0.55%, 3.55%, and 1.75% over SAGAT-BERT, demonstrating strong cross-domain robustness.
BibTeX
@inproceedings{wang2025dual,
  author = {Xinyu Wang and Yuhang Gao and Xiao Lv and Kun Zhong and Peng Hui Tan and Zhaobin Tan},
  title = {{Dual-Tree-Guided Aspect-Based Sentiment Analysis Incorporating Structure-Aware Semantic Refinement and Graph Attention}},
  booktitle = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17091492},
}
Figure